A recommendation model-oriented adaptive federated learning optimization method and system

Abstract

The application provides a kind of adaptive joint learning optimization method and system for recommendation model, the method of the application includes: obtaining the training data set of recommendation model;According to the sampling strategy of divergent goods, training data set is sampled to obtain training sample containing positive sample, negative sample and similar sample;The joint loss function containing point state loss, homogeneity loss and heterogeneity loss is constructed, the importance weight of point state loss, homogeneity loss and heterogeneity loss is dynamically calculated according to real-time prediction error in training process, and point state loss, homogeneity loss and heterogeneity loss are weighted and fused using importance weight;Training sample is input into recommendation model, and the model parameters of recommendation model are iteratively updated using the joint loss function after weighted fusion, to obtain trained recommendation model;The user and candidate item are scored and predicted using the trained recommendation model, and the optimal recommendation list is generated according to the predicted score.

Description

Technical Field

[0001] This invention relates to the field of recommender system technology, and in particular to an adaptive joint learning optimization method and system for recommender models. Background Technology

[0002] Recommender systems aim to learn user preferences from past interactions, helping users quickly select their favorite items from a vast array of products. Early recommendation methods relied on a point-state objective, minimizing the error between model predictions and actual values. However, since systems typically only observe positive feedback data (e.g., clicks), negative sample sampling methods are needed, sampling a portion of the set of items the user hasn't interacted with as negative samples. This often leads to system bias. Therefore, researchers have proposed a pairwise ranking method using a heterogeneity loss function. This method optimizes the model by maximizing the difference between positive and negative samples, avoiding the influence of directly predicting negative sample labels.

[0003] With the development of online services, explicit feedback information such as ratings and viewing time has become increasingly abundant. Researchers have further developed pairwise ranking methods based on explicit feedback, which utilize explicit feedback to better capture user preferences. Recent studies have shown that explicit feedback and implicit data share a latent subspace that satisfies the characteristics of both types of data and exhibits good generalization. Subsequently, researchers proposed an alternating pairwise ranking optimization method that simultaneously optimizes both the pairwise ranking objective and the point-state objective. Subsequent improvements have further explored the partial order relationships or collaborative information in explicit feedback and implicit data. However, these methods only control the influence of different objectives with predefined hyperparameters, preventing the learning process from dynamically adjusting to different users. Furthermore, although some methods have combined point-state objectives and heterogeneity losses, they merely mechanically integrate these two losses without deeply modeling the homogeneity of items.

[0004] Modeling homogeneity can mitigate the impact of imbalanced data distribution. For heterogeneity loss, imbalanced data distributions excessively amplify the gap between positive and negative samples, causing the learned intervals to exceed a reasonable range and deviate from the user's true preference differences. This excessive separation distorts the relative preference relationships between items, ultimately reducing the model's ability to recover the user's true ranking behavior. Modeling homogeneity, on the other hand, can help narrow the score differences between niche and popular items, thereby meeting users' personalized needs.

[0005] However, simultaneously optimizing point-state objectives, homogeneous losses, and heterogeneous losses presents challenges—their optimization directions are not always aligned. Because each emphasizes different objectives, gradients of different objectives in the parameter space may interfere with or cancel each other out, leading to training instability or slow convergence. Furthermore, if the weights of these three objectives are not properly balanced, the model may get stuck in local optima, or experience overfitting or underfitting. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive joint learning optimization method and system for recommendation models, aiming to solve the above-mentioned problems in the prior art.

[0007] This invention provides an adaptive joint learning optimization method for recommendation models, comprising: Obtain the training dataset for the recommendation model; wherein, the training dataset includes historical interaction records between users and items; Training samples containing positive, negative, and similar samples are obtained from the training dataset according to the divergent item sampling strategy; A joint loss function comprising point state loss, homogeneity loss, and heterogeneity loss is constructed. During training, the importance weights of the point state loss, homogeneity loss, and heterogeneity loss are dynamically calculated based on the real-time prediction error. The point state loss, homogeneity loss, and heterogeneity loss are then weighted and fused using the importance weights. The training samples are input into the recommendation model, and the model parameters of the recommendation model are iteratively updated using the weighted fusion joint loss function to obtain the trained recommendation model. The trained recommendation model is used to predict user and candidate item ratings, and an optimal recommendation list is generated based on the predicted scores.

[0008] This invention provides an adaptive joint learning optimization system for recommendation models, comprising: The data acquisition module is used to acquire the training dataset for the recommendation model; wherein, the training dataset includes historical interaction records between users and items; The divergence sampling module is used to sample training samples containing positive samples, negative samples, and similar samples from the training dataset according to the divergence item sampling strategy. The loss construction and weighting module is used to construct a joint loss function that includes point state loss, homogeneity loss and heterogeneity loss. During training, the importance weights of the point state loss, homogeneity loss and heterogeneity loss are dynamically calculated according to the real-time prediction error, and the point state loss, homogeneity loss and heterogeneity loss are weighted and fused using the importance weights. The model training module is used to input the training samples into the recommendation model and iteratively update the model parameters of the recommendation model using the weighted fusion joint loss function to obtain the trained recommendation model. The recommendation generation module is used to predict the ratings of users and candidate items using the trained recommendation model, and generate an optimal recommendation list based on the predicted scores.

[0009] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above-described adaptive joint learning optimization method for recommendation models.

[0010] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described adaptive joint learning optimization method for recommendation models.

[0011] The embodiments of this invention can include the following beneficial effects: This invention proposes an Adaptive Joint Learning (AJL) framework for recommendation systems, which dynamically balances the influence of point states, homogeneity, and heterogeneity objectives in joint learning. Specifically, AJL trains the recommendation model by jointly optimizing these three objective functions. The weights of the objective functions are allocated by a weighting method based on real-time error, which dynamically adjusts the weights by mapping the real-time training error to a uniform scale and converting it into importance weights, thereby balancing these objectives. Furthermore, to further explore the homogeneity and heterogeneity among items, this invention proposes divergent item pairs, i.e., item pairs for which different users have differing attitudes. Although item pairs contain richer information on homogeneity and heterogeneity, existing methods struggle to effectively align the representations of these items for learning. Therefore, this invention proposes a divergent item sampling strategy to identify divergent item pairs and extract training samples from these items, guiding the model to learn more accurate user preferences rather than overfitting the data distribution. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the adaptive joint learning optimization method for recommendation models according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of an adaptive joint learning optimization system for recommendation models according to an embodiment of the present invention. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0015] Method Implementation Examples According to embodiments of the present invention, an adaptive joint learning optimization method for recommendation models is provided. Figure 1 This is a flowchart of the adaptive joint learning optimization method for recommendation models according to an embodiment of the present invention, as follows: Figure 1 As shown, the adaptive joint learning optimization method for recommendation models according to an embodiment of the present invention specifically includes: Step S101: Obtain the training dataset for the recommendation model; wherein, the training dataset includes historical interaction records between users and items; the recommendation model is a matrix factorization model, expressed as: Formula 7; in, This represents the predicted score for item i by user u. The vector representation of user u, This represents the user matrix composed of all user vectors. express The Middle Line number Column elements; This represents an item matrix consisting of all item vectors. Representation matrix The OK, Represents an item vector The k-th dimension, This indicates the preset vector dimension.

[0016] Step S102 involves sampling training samples containing positive, negative, and similar samples from the training dataset according to the divergent item sampling strategy; specifically including: Calculate the positive probability, equal probability, and negative probability between item pairs based on the comparison of user ratings for item pairs; Based on the comparison between the positive probability and the equal probability and negative probability, positive and negative item pairs are sampled to obtain positive samples and negative samples; Based on the similarity between positive samples and candidate similar items, and the similarity ratio between positive and negative samples and candidate similar items, similar item pairs are sampled to obtain similar samples; The formula for calculating the positive probability is: Formula 2; The formula for calculating the equal probability is: Formula 3; The formula for calculating the negative probability is: Formula 4; in, This indicates that for items in the entire dataset The rating is greater than the The probability of a rating; This indicates that for items in the entire dataset The rating is equal to the The probability of a rating; This indicates that for items in the entire dataset The rating is less than that of The probability of a rating; Represents the entire set of users; This represents the rating that user u gives to item i. This represents the rating that user u gives to item j.

[0017] Step S103: Construct a joint loss function comprising point-state loss, homogeneity loss, and heterogeneity loss. During training, dynamically calculate the importance weights of each of the point-state loss, homogeneity loss, and heterogeneity loss based on the real-time prediction error, and then use these importance weights to perform a weighted fusion of the point-state loss, homogeneity loss, and heterogeneity loss; specifically including: Construct the joint loss function as shown in Equation 1; Formula 1; in, , and Let these represent homogeneity loss, heterogeneity loss, and point state loss, respectively. , , These represent the corresponding weights. Represents model parameters The regularization coefficient; During training, the global features, user features, item features, and real-time prediction error of the training dataset are obtained. The global features, user features, item features, and real-time prediction error are weighted and combined according to the number of training iterations to calculate the importance weights of heterogeneity loss, homogeneity loss, and point state loss respectively. The importance weights are normalized using the softmax function to obtain normalized weight values. The heterogeneity loss, homogeneity loss, and point state loss are weighted and summed using the normalized weight values ​​to obtain the weighted fusion joint loss function; The global features include PDI and EDI; PDI represents the probability of inconsistencies in ratings between item pairs, and its expression is: Formula 5; The EDI represents the standard deviation of the user rating standard deviation, expressed as: Formula 6; Where I represents the set of items and U represents the set of users. This represents the probability that item i has a higher rating than item j. This represents the standard deviation function.

[0018] Step S104 involves inputting the training samples into the recommendation model and iteratively updating the model parameters using the weighted fusion joint loss function to obtain the trained recommendation model; specifically including: S41: Input the user information and item information from the training samples into the recommendation model to obtain the predicted score; S42: Substitute the predicted score and the real score in the training sample into the weighted fusion joint loss function to calculate the current loss value; S43: Based on the current loss value, calculate the gradient of the model parameters of the recommendation model using the backpropagation algorithm; S44: Update the model parameters using the optimizer based on the gradient; S45: Repeat steps S41 to S44 until the preset number of training rounds is reached or the preset convergence condition is met.

[0019] Step S105: Use the trained recommendation model to predict the ratings of users and candidate items, and generate the optimal recommendation list based on the predicted scores.

[0020] The following details the adaptive joint learning optimization method for recommendation models according to embodiments of the present invention, such as... Figure 2As shown, the above technical solutions of the embodiments of the present invention will be described in detail.

[0021] The specific steps of this embodiment of the invention are as follows: Step 1: Divergent Item Sampling Strategy Due to the highly uneven distribution of divergent item pairs and the inherent characteristics of pairwise ranking methods, it is difficult to recommend niche items. Joint optimization methods often exaggerate the gap between positive and negative samples, deviating from the true preference difference; on the other hand, SPR (Similarity-based Pairwise Ranking) is highly sensitive to the distribution of divergent item pairs, and may get stuck in trivial solutions when divergent item pairs are evenly distributed, such as assigning almost the same score to different items through similarity constraints.

[0022] To address these issues, embodiments of the present invention consider two strategies: A. reducing the probability of popular items being positive samples and niche items being negative samples; B. reducing the likelihood of samples similar to both positive and negative samples being classified as similar samples. Based on these strategies, embodiments of the present invention propose a divergent item sampling strategy to construct more informative training data.

[0023] 1. Item Relationship Analysis For users and two items and There are three types of relationships between these two items, namely (This indicates that a user u has a relationship with an item) The rating is greater than the (ratings) (This indicates that a user u has a relationship with an item) The rating is equal to the (ratings) and (A user u has a pair of items) The rating is less than that of (Ratings). (Using...) , and This represents the probability of these three relationships in the entire dataset. This indicates the prevalence of this item in the dataset, assuming here... For item pairs ,have: (1); Similarly, we can obtain and : (2); (3); in, This indicates that in the entire dataset, items The rating is greater than the The probability of a rating; This indicates that in the entire dataset, items The rating is equal to the The probability of a rating; This indicates that in the entire dataset, items The rating is less than that of The probability of a rating; Represents the entire set of users; This represents the rating that user u gives to item i.

[0024] 2. Negative sample sampling According to strategy A, this embodiment of the invention focuses on item pairs with lower probabilities. Therefore, for a positive-negative item pair, its sampling probability is... and Proportional to, with Inversely proportional. Assume the user's maximum rating for an item is... The minimum rating is 1; user historical data is used It can be broken down into a set, namely For each set ,have: (4); Then, for each group Each item in Perform iterations, treating it as a positive sample and from its neighbor set Select some items as negative samples to match with the set. Each item in The probability of being selected as a negative sample is given by the following formula: (5); (6); in It is a smoothing constant used to smooth out the influence of the probability of the segmentation terms; Indicates positive samples Under the premise of negative samples Let i be the probability of an item i. This represents another item that is different from item i; This indicates that any item i is selected as a positive sample. The probability score; This represents the sigmoid function, i.e., y = 1 / 1 + exp(-x).

[0025] Therefore, for any pair of items The probability that they form a positive-negative sample pair is (After excluding the effect of the smoothing factor). When hour, Numerically, it approximates a linear function; therefore, the probability of two items being selected as positive or negative samples is almost identical, and the difference between them is not significant. However, when hour, The function gradually deviates from the linear function and remains below the line, indicating that for popular items, the function effectively weakens their dominance, thus preventing the gap between popular and niche samples from being overemphasized and promoting a more balanced distribution.

[0026] 3. Sampling of similar items When sampling similar samples, two factors should be considered: (a) the similarity between the positive sample and the target item. The more similar the target item is to the current item, the greater the likelihood that it will be selected as a similar item. (b) the similarity ratio between the positive sample, negative sample, and target item. If the negative sample has a higher similarity to the target item, then the probability of the target item being a similar item to the positive sample should decrease. Therefore, from... Sampling similar terms, for each The probability that it is sampled as a similar item is given by the following formula: (7); (8); in, Indicates positive samples Under the premise of similar samples The probability of an item i. Indicates positive samples and Similar scores, with a smoothing factor; Indicates positive samples and negative samples Under the premise of choosing As a score for similar items; This indicates that in the entire dataset, items The rating is equal to the The probability of a rating; This indicates that in the entire dataset, items The rating is equal to the The probability of a rating.

[0027] Compared with uniform sampling, Equation (8) tends to select items that are similar to positive samples but not similar to negative samples.

[0028] Step 2: Build a recommendation prediction model (9); The constructed recommendation prediction model is shown in formula (9), where, This represents the predicted score for item i by user u. The vector representation of user u, This represents the user matrix composed of all user vectors. It is a matrix The OK, yes The Middle Line number Column elements, This represents an item matrix consisting of all item vectors. Represents any item The vector representation of a matrix is ​​also a matrix. The OK; This indicates the preset vector dimension.

[0029] Step 3: Training and Optimization of the Recommendation Model Based on the Joint Loss Function Point-like objects, homogeneity, and heterogeneity are fundamental characteristics of objects and play a crucial role in modeling user-object interactions. Although point-like objects encompass both homogeneity and heterogeneity, their effectiveness is influenced by the backbone model trained on the view, leading to suboptimal results in practical applications. Some methods have explored the benefits of alternatively learning point-like objects and heterogeneity, while others have demonstrated the advantages of jointly learning homogeneity and heterogeneity. However, none of these methods adequately consider the relationships between objects, resulting in suboptimal performance in certain situations.

[0030] Therefore, the goal of this invention is to jointly optimize these objectives, which brings complementary advantages: point states ensure that the model conforms to observed user feedback, heterogeneity captures the relative preference order between positive and negative items, and homogeneity enforces consistency among similar items to avoid over-amplifying preference differences. Based on these advantages, the AJL proposed in this invention utilizes all three simultaneously and dynamically calculates their weights using a weighting method based on real-time error.

[0031] 1. Objective function The system optimizes the proposed recommendation model using formula (10) to obtain the optimal vector representations of users and items: (10); In the formula, , and These represent homogeneous, heterogeneous, and point-state targets, respectively. , and These are the hyperparameters used to control the effects of these three characteristics; Represents model parameters The regularization coefficient; D represents the dataset used for training, r with a hat indicates the model's prediction result, and r without a hat indicates the true label in the dataset; This represents the model's prediction of user u for items. The rating; This represents the model's prediction of user u for items. The rating; This represents the model's prediction of user u for items. The rating; This represents the actual rating of user u for item j recorded in the dataset; This represents the model's predicted rating of user u for item j; the range of values ​​for item j is... .

[0032] 2. Dynamic weighting based on real-time error Equation (10) involves three hyperparameters to control the contributions of point-state, homogeneous, and heterogeneous objectives. A common method for selecting these hyperparameters is grid search; however, grid search is computationally expensive and can only yield a single global configuration, failing to adapt to changes in the importance of different objectives across different training samples or training stages. More importantly, each objective plays a different role at different stages of model convergence. Based on this observation, this embodiment of the invention proposes dynamically determining the weights of different objectives during training based on real-time prediction errors.

[0033] Specifically, in order to model the importance of point state loss, heterogeneity loss, and homogeneity loss, embodiments of the present invention explore the following features: (1) Prior weights, including global features of the dataset User characteristics and the characteristics of the items ; (2) Real-time weights, including real-time error and number of training iterations .

[0034] Finally, the weights of each loss function are defined as follows: (11); For global features, this invention proposes two metrics: and .

[0035] The situation where two items have inconsistent ratings in the dataset; (12); Standard deviation of user rating standard deviation; (13); therefore, , , Superscript Indicates heterogeneity loss. Indicates homogeneous loss. This represents the point state loss.

[0036] Furthermore, PDI and EDI can be extended to both the user and item sides to calculate user characteristics and item characteristics.

[0037] (14); (15); (16); During training, the model should assign greater weights to larger errors. Since different loss functions have different optimization objectives and therefore different units of measurement, it is necessary to unify the units of measurement. Here, the errors in modeling different features are converted into importance, and then used... Normalize: Point state characteristics: ; Homogeneity: ; Heterogeneity: ; in addition, The number of iterations affects the impact of different losses as the number of training iterations increases. For point-state loss, it is very important in the early stages, as it can quickly approximate the true value; in the later stages, the model gradually converges, and the learning effect is poor.

[0038] For heterogeneous loss, the early stage is important as it can quickly distinguish between positive and negative samples; the later stage will amplify the influence of data distribution and deviate from the true result.

[0039] For homogeneous loss, the early learning results are not accurate enough, so its early impact is limited; however, it can effectively mitigate the harm of heterogeneous loss in the later stages.

[0040] Therefore, for point state loss and heterogeneous loss, For homogeneous losses, .

[0041] Global features of the comprehensive dataset User characteristics and the characteristics of the items Real-time error Given the number of training iterations t, the weights of the heterogeneity loss function are: (17); The weights of the homogeneity loss function are: (18); The weights of the point-state loss function are: (19); Finally, the application The function is used to obtain the final weights of each objective function for each training sample: (20).

[0042] Step 4: Item Score Prediction and Sorting Based on the optimal vector representation of each user and the optimal vector representation of each item obtained in step 3, the scoring function is used to predict the user's preference score for each item. Then, all items are sorted according to the scores. Finally, a recommendation list is generated according to actual needs. The scoring function is shown in formula (9).

[0043] In summary, this invention proposes an AJL (Advanced Handly Selection Language) that dynamically balances the impact of these three objectives by mapping real-time training errors to importance weights. It also proposes a segmented item sampling strategy to extract training samples, guiding the model to learn more accurate user preferences rather than overfitting the data distribution. Experimental results on four datasets demonstrate that the proposed method outperforms state-of-the-art methods, achieving an average improvement of 20.47% on Precision@5 and an average improvement of 7.36% on NDCG@5.

[0044] System Implementation Examples According to embodiments of the present invention, an adaptive joint learning optimization system for recommendation models is provided. Figure 3 This is a schematic diagram of an adaptive joint learning optimization system for recommendation models according to an embodiment of the present invention, as shown below. Figure 3 As shown, the adaptive joint learning optimization system for recommendation models according to an embodiment of the present invention specifically includes: The data acquisition module 30 is used to acquire the training dataset of the recommendation model; wherein, the training dataset includes historical interaction records between users and items; The divergence sampling module 32 is used to sample training samples containing positive samples, negative samples and similar samples from the training dataset according to the divergence item sampling strategy; The loss construction and weighting module 34 is used to construct a joint loss function that includes point state loss, homogeneity loss and heterogeneity loss. During the training process, the importance weights of the point state loss, homogeneity loss and heterogeneity loss are dynamically calculated according to the real-time prediction error, and the point state loss, homogeneity loss and heterogeneity loss are weighted and fused using the importance weights. The model training module 36 is used to input the training samples into the recommendation model and iteratively update the model parameters of the recommendation model using the weighted fusion joint loss function to obtain the trained recommendation model. The recommendation generation module 38 is used to predict the ratings of users and candidate items using the trained recommendation model, and generate an optimal recommendation list based on the predicted scores.

[0045] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0046] Device Example 1 This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it performs the steps described in the method embodiment.

[0047] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.

[0048] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive joint learning optimization method for recommendation models, characterized in that, include: Obtain the training dataset for the recommendation model; wherein, the training dataset includes historical interaction records between users and items; Training samples containing positive, negative, and similar samples are obtained from the training dataset according to the divergent item sampling strategy; A joint loss function comprising point state loss, homogeneity loss, and heterogeneity loss is constructed. During training, the importance weights of the point state loss, homogeneity loss, and heterogeneity loss are dynamically calculated based on the real-time prediction error. The point state loss, homogeneity loss, and heterogeneity loss are then weighted and fused using the importance weights. The training samples are input into the recommendation model, and the model parameters of the recommendation model are iteratively updated using the weighted fusion joint loss function to obtain the trained recommendation model. The trained recommendation model is used to predict user and candidate item ratings, and an optimal recommendation list is generated based on the predicted scores.

2. The method according to claim 1, characterized in that, The training samples, which include positive samples, negative samples, and similar samples, are specifically obtained from the training dataset according to the divergent item sampling strategy. Calculate the positive probability, equal probability, and negative probability between item pairs based on the comparison of user ratings for item pairs; Based on the comparison between the positive probability and the equal probability and negative probability, positive and negative item pairs are sampled to obtain positive samples and negative samples; Based on the similarity between positive samples and candidate similar items, and the similarity ratio between positive and negative samples and candidate similar items, similar item pairs are sampled to obtain similar samples.

3. The method according to claim 1, characterized in that, A joint loss function is constructed, comprising point-state loss, homogeneity loss, and heterogeneity loss. During training, the importance weights of each of the point-state loss, homogeneity loss, and heterogeneity loss are dynamically calculated based on the real-time prediction error. The point-state loss, homogeneity loss, and heterogeneity loss are then weighted and fused using these importance weights. Specifically, this includes: Construct the joint loss function as shown in Equation 1; Official 1; in, , and Let these represent homogeneity loss, heterogeneity loss, and point state loss, respectively. , , These represent the corresponding weights. Represents model parameters The regularization coefficient; During training, the global features, user features, item features, and real-time prediction error of the training dataset are obtained. The global features, user features, item features, and real-time prediction error are weighted and combined according to the number of training iterations to calculate the importance weights of heterogeneity loss, homogeneity loss, and point state loss respectively. The importance weights are normalized using the softmax function to obtain normalized weight values. The heterogeneity loss, homogeneity loss, and point state loss are weighted and summed using the normalized weight values ​​to obtain the weighted fusion joint loss function.

4. The method according to claim 1, characterized in that, The training samples are input into the recommendation model, and the model parameters of the recommendation model are iteratively updated using the weighted fusion joint loss function to obtain the trained recommendation model, specifically including: S41: Input the user information and item information from the training samples into the recommendation model to obtain the predicted score; S42: Substitute the predicted score and the real score in the training sample into the weighted fusion joint loss function to calculate the current loss value; S43: Based on the current loss value, calculate the gradient of the model parameters of the recommendation model using the backpropagation algorithm; S44: Update the model parameters using the optimizer based on the gradient; S45: Repeat steps S41 to S44 until the preset number of training rounds is reached or the preset convergence condition is met.

5. The method according to claim 2, characterized in that, The formula for calculating the positive probability is: Official 2; The formula for calculating the equal probability is: Official 3; The formula for calculating the negative probability is: Official 4; in, This indicates that for items in the entire dataset The rating is greater than the The probability of a rating; This indicates that for items in the entire dataset The rating is equal to the The probability of a rating; This indicates that for items in the entire dataset The rating is less than that of The probability of a rating; Represents the entire set of users; This represents the rating that user u gives to item i. This represents the rating that user u gives to item j.

6. The method according to claim 3, characterized in that, The global features include PDI and EDI; wherein, PDI represents the probability of inconsistency in ratings between item pairs, and is expressed as: Official 5; The EDI represents the standard deviation of the user rating standard deviation, expressed as: Official 6; Where I represents the set of items and U represents the set of users. This represents the probability that item i has a higher rating than item j. This represents the standard deviation function.

7. The method according to claim 1, characterized in that, The recommendation model is a matrix factorization model, and its expression is: Official 7; in, This represents the predicted score for item i by user u. The vector representation of user u, This represents the user matrix composed of all user vectors. express The Middle Line number Column elements; This represents an item matrix consisting of all item vectors. Representation matrix The OK, Represents an item vector The k-th dimension, This indicates the preset vector dimension.

8. An adaptive joint learning optimization system for recommendation models, characterized in that, include: The data acquisition module is used to acquire the training dataset for the recommendation model; wherein, the training dataset includes historical interaction records between users and items; The divergence sampling module is used to sample training samples containing positive samples, negative samples, and similar samples from the training dataset according to the divergence item sampling strategy. The loss construction and weighting module is used to construct a joint loss function that includes point state loss, homogeneity loss and heterogeneity loss. During training, the importance weights of the point state loss, homogeneity loss and heterogeneity loss are dynamically calculated according to the real-time prediction error, and the point state loss, homogeneity loss and heterogeneity loss are weighted and fused using the importance weights. The model training module is used to input the training samples into the recommendation model and iteratively update the model parameters of the recommendation model using the weighted fusion joint loss function to obtain the trained recommendation model. The recommendation generation module is used to predict the ratings of users and candidate items using the trained recommendation model, and generate an optimal recommendation list based on the predicted scores.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the adaptive joint learning optimization method for recommendation models as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the adaptive joint learning optimization method for recommendation models as described in any one of claims 1-7.

Images