Project recommendation method and system based on causal popularity debiasing

Abstract

The application provides a project recommendation method and system based on causal popularity debiasing, and the method comprises the following steps: obtaining fixed project popularity corresponding to each project based on the project popularity of each exposed project in each historical time period; obtaining a predicted score between a user and an exposed project based on a pre-stored recommendation model, the fixed project popularity, an optimized user embedding vector set and an optimized project embedding vector set; and recommending the exposed project to the user based on the predicted score. The application can alleviate the influence of popularity bias on the recommendation result by fusing the project popularity and the user interest to quantify the user consistency.

Description

Technical Field

[0001] This invention relates to the field of recommendation system technology, and in particular to a method and system for item recommendation based on causal popularity debiasing. Background Technology

[0002] Recommender systems leverage users' historical behavior to uncover the matching relationships between user interests and item attributes. Based on predicted matching scores, they generate a list of recommended items that each user might be interested in, thus providing personalized recommendations quickly and accurately. However, due to popularity bias, a few popular items in the recommendation system will become increasingly popular and attract more user attention, while most unpopular items will become increasingly difficult to expose, leading to the "Matthew effect" in recommender systems.

[0003] Currently available model frameworks for popularity bias removal include: Model-Agnostic Counterfactual Reasoning (MACR) framework, which constructs a causal graph of the recommendation process (edges pointing from user and item variables to the recommendation result represent the negative impact of item popularity on the recommendation result), and uses counterfactual inference to treat the recommendation result affected by popularity as a hyperparameter, removing it when predicting the matching score; Popularity-bias deconfounding and adjusting (PDA) framework, which proposes that the impact of item popularity is two-sided, uses causal intervention to remove adverse effects during the training phase and utilizes its positive effects during the inference phase, significantly improving the matching degree between user interests and item attributes; and the Disentangling Interest and Conformity with Causal Embedding (DICE) framework, which takes a user-centric approach, arguing that the factors determining popularity bias are user interests and user conformity. Conformity (i.e., user herd mentality) was considered, and an embedded representation framework with specific causal relationships to decouple popularity bias was designed. The time-aware disentangled framework (TIDE) attributes the cause of popularity bias to item quality and user consistency, with the former having a positive impact and the latter being considered a negative impact. Utilizing time information, this framework generates clicks from three parts: static item quality, dynamic consistency effect, and user-item matching score returned by any recommendation model. Finally, intervention inference is performed so that recommendations can benefit from benign popularity bias while avoiding harmful bias.

[0004] However, existing model frameworks for popularity bias correction fail to accurately reflect the impact of user consistency on recommendation results. The MACR framework considers user consistency from the user's perspective and views it as a negative impact on recommendation results; the PDA framework considers user consistency as a positive impact of popularity bias and does not model or quantify user consistency; the DICE framework divides the training set into two categories: user interests and user consistency, and establishes two classes of embedding vector representations for users and items respectively, calculating the score for each class through dot product; the TIDE framework explicitly models user consistency in a causal graph, using specific parameters to accumulate interaction feedback from past times and reducing the contribution of past times based on time intervals.

[0005] Therefore, there is an urgent need for an item recommendation method that can truly reflect the impact of user consistency on recommendation results, in order to reduce the impact of popularity bias on the recommendation system. Summary of the Invention

[0006] In view of this, embodiments of the present invention provide a method and system for item recommendation based on causal popularity debiasing. The method models user consistency based on the negative impact of item popularity on user interests, quantifies user consistency by fusing item popularity and user interests, and balances it with recommendation objectives during the training phase, thereby mitigating the impact of user consistency in the recommendation system.

[0007] One aspect of the present invention provides a method for item recommendation based on causal popularity debiasing, the method comprising the following steps:

[0008] Based on the project popularity of each exposed project in various historical time periods, a fixed project popularity is obtained for each project.

[0009] Based on the pre-stored recommendation model, fixed item popularity, optimized user embedding vector set, and optimized item embedding vector set, a predicted score between the user and the exposed items is obtained, and exposed items are recommended to the user based on the predicted score;

[0010] The optimized user embedding vector and optimized item embedding vector are obtained based on the user embedding vectors corresponding to users in the training set, the item embedding vectors corresponding to exposed items, and the item popularity of each exposed item across multiple historical time periods, through the following optimization methods:

[0011] A matching score is obtained based on any item embedding vector, the popularity of the item corresponding to the item embedding vector, any user embedding vector, and a pre-stored recommendation model. A consistency score is obtained based on the item popularity, the user embedding vector, and a pre-stored user consistency model.

[0012] The training objective function is obtained based on the matching score and consistency score, and the user embedding vector and item embedding vector are iteratively optimized based on the training objective function.

[0013] In some embodiments of the present invention, the user embedding vector is obtained by vectorizing user data based on embedding techniques;

[0014] Project embedding vectors are obtained by vectorizing the data of exposed projects based on embedding techniques.

[0015] In some embodiments of the present invention, a matching score is obtained based on any item embedding vector, the item popularity corresponding to that item embedding vector, any user embedding vector, and a pre-stored recommendation model, including:

[0016] Input any user embedding vector and item embedding vector into a pre-stored recommendation model, and perform a dot product operation between the model output and the item popularity to obtain a matching score.

[0017] A consistency score is obtained based on the project's popularity, the user's embedding vector, and a pre-stored user consistency model, including:

[0018] The project popularity and user embedding vectors are input into a pre-stored user consistency model, and the user embedding vector is calculated by dividing the sum of the project popularity and a set constant, and this value is used as the consistency score.

[0019] In some embodiments of the present invention, the training objective function is obtained based on the matching score and the consistency score, including:

[0020] Based on the matching scores obtained from positive and negative samples, a recommendation loss function is constructed using a Bayesian personalized ranking algorithm; where positive samples are items that have historical interactions with the user, and negative samples are items that have not historical interactions with the user.

[0021] The consistency loss function is derived based on the binary cross-entropy loss function and the consistency score.

[0022] The training objective function is obtained based on the consistency loss function, the recommendation loss function, and the setting of hyperparameters.

[0023] In some embodiments of the present invention, iterative optimization of user embedding vectors and item embedding vectors based on a training objective function includes:

[0024] The adaptive moment estimation algorithm is used to iteratively optimize the feature values ​​in the user embedding vector and the item embedding vector based on the training objective function.

[0025] In some embodiments of the present invention, a fixed project popularity is obtained for each project based on the project popularity of each exposed project in various historical time periods, including:

[0026] Based on the popularity of each exposed project across multiple historical time periods, calculate the time context weight coefficient for each historical time period.

[0027] For each exposed project, the project popularity and time context weight coefficient for each historical time period are multiplied together, and the products are summed to obtain the fixed project popularity for each project.

[0028] In some embodiments of the present invention, based on the project popularity of each exposed project across multiple historical time periods, a time context weight coefficient corresponding to each historical time period is calculated, including:

[0029] Calculate the mean popularity of all items across all historical time periods, and use the ratio of the mean popularity of items in each historical time period to the sum of the mean popularity of items across all historical time periods as the temporal context weight coefficient for each historical time period; or

[0030] Calculate the difference between the mean project popularity of each historical time period and the mean project popularity of all projects in the previous historical time period. Use the ratio of the difference between the mean project popularity of each historical time period and the sum of the differences between the mean project popularity of all historical time periods as the time context weight coefficient of each historical time period.

[0031] In some embodiments of the present invention, a predicted score between a user and exposed items is obtained based on a pre-stored recommendation model, fixed item popularity, an optimized user embedding vector set, and an optimized item embedding vector set, including:

[0032] Select any optimized user embedding vector and optimized item embedding vector from the optimized user embedding vector set and the optimized item embedding vector set, input them into the pre-stored recommendation model, and perform a dot product operation between the model result and the corresponding fixed item popularity to obtain the predicted scores between users and items corresponding to the selected optimized user embedding vector and optimized item embedding vector, respectively.

[0033] Repeat the above steps to obtain the predicted scores between all users and items corresponding to the optimized user embedding vector set and the optimized item embedding vector set;

[0034] Recommending exposed items to users based on predicted scores includes: sorting the recommended items for users in descending order of predicted scores, based on the optimized item embedding vectors.

[0035] Another aspect of the present invention provides an item recommendation system based on causal popularity debiasing, including a processor, a memory, and a computer program / instructions stored in the memory. The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method described in any of the above embodiments.

[0036] Another aspect of the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0037] This invention proposes a project recommendation method and system based on causal popularity bias correction. During the training phase, it intervenes in projects to eliminate the interference of confounding factors such as project popularity. Furthermore, it models user consistency by fusing project popularity with user embedding vectors through quantification, thereby mitigating user consistency issues. This invention can alleviate the impact of popularity bias on recommendation results from the perspectives of both users and projects, enabling the recommendation system to learn more realistic and accurate user interests.

[0038] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0039] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0040] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0041] Figure 1 This is an example diagram of a cause-effect graph in one embodiment of the present invention.

[0042] Figure 2 This is a recommendation causal graph of an item recommendation method based on causal popularity debiasing in one embodiment of the present invention.

[0043] Figure 3 This is a flowchart illustrating the optimization operation during the training phase of the project recommendation method based on causal popularity debiasing in one embodiment of the present invention.

[0044] Figure 4 This is a schematic diagram of a time-aware causal popularity-based bias-removal recommendation algorithm framework in one embodiment of the present invention.

[0045] Figure 5 This is a flowchart illustrating the inference phase of a project recommendation method based on causal popularity debiasing in one embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0047] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0048] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0049] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0050] A cause-effect graph is a directed acyclic graph, which can be represented by G = {V, E}, where V is the set of variables, i.e., the nodes in the cause-effect graph; and E is the causal relationship between the variables, i.e., the directed edges in the cause-effect graph. For example... Figure 1 As shown, X→Y represents the causal effect between variables X and Y, while variable Z is a confounding factor of this causal effect because of information interference from the non-causal direction X←Z→Y.

[0051] To remove the influence of the confounding factor Z, physical intervention experiments or randomized controlled trials can be conducted, but these are difficult to implement in practice. The do operator can be used to eliminate this confounding factor. Figure 1 All arrows pointing to X are used to prevent any information related to X from flowing in non-causal directions. The steps are as follows: after identifying the variables that are confounding factors in the causal graph, the confounding factors are statistically adjusted according to the do calculus rules to achieve deconfounding using observation data and estimate the true causal effect.

[0052] Assuming users can only interact with items already displayed in the recommendation system, the recommendation problem can be simplified to finding the matching relationship between user interests and item attributes. In the recommendation system, U = {u1, u2, ..., u...} n} can represent the user set, where n represents the number of users in the recommendation system; I = {i1, i2, ..., i m} can represent the set of items that have been exposed in the recommendation system, and m represents the number of items; This can be used to represent a user-item interaction matrix, where the interaction degree y is calculated if user u interacts with item i. ui This can be denoted as 1; if there is no interaction between user u and item i, then the interaction degree y is 1. ui This can be represented as 0 (the representation of interaction degree mentioned in the embodiments of this invention is only an example, and this invention does not specifically limit it). Based on the user-item interaction matrix, the recommendation algorithm learns the matching relationship f(u,i│θ) between user interests and item attributes (θ is the relevant learning parameter), and then predicts the degree of interest of user u and item i. Here, user interest refers to the degree of interest of the user in the item, and item attributes refer to the item tags that the user can see, such as the item's price, brand, manufacturer, and reviews.

[0053] Existing model frameworks for popularity bias removal suffer from drawbacks such as ignoring the impact of time-varying factors and being one-sided in terms of user consistency. To address this, this invention provides a project recommendation method based on causal popularity bias removal and establishes a time-aware causal graph for recommendation. During the project recommendation process, user interests and project attributes can be represented in the form of embedded vectors.

[0054] like Figure 2 As shown, the training and recommendation phases of the item recommendation method based on causal popularity bias proposed in this invention can be explained using a causal graph to illustrate the training and inference processes of the recommendation system undergoing popularity bias bias removal. Figure 2 The causal graph includes five variables and their causal relationships: user, project, project popularity, time, and recommendation results (divided into two categories: recommended and not recommended). Figure 2 (a) in the figure is a time-aware recommendation causal graph with popularity bias, in which item popularity has a causal relationship with time, item popularity serves as a confounding factor between items and recommendation results, and users, items, and item popularity can all promote the generation of recommendation results. Therefore, this invention explains the main reasons for popularity bias from the perspective of item popularity and time.

[0055] Figure 2(b) in the diagram represents the causal graph of a recommendation system with popularity bias during the training phase. Training yields optimized user embedding vector sets and optimized item embedding vector sets corresponding to users and items. First, due to user consistency, users are easily influenced by item popularity, deviating from their true interests. This user consistency leads to a negative impact of popularity bias from the user's perspective, reducing the accuracy of the recommendation results. To address the modeling problem of user consistency, a quantification method integrating item popularity and user embedding vectors is proposed. By balancing user consistency with recommendation accuracy, the influence of user consistency on the recommendation results is removed. Second, to address the confounding problem of item popularity, an intervention method is proposed. This involves controlling popularity to achieve deconfounding and mitigate the negative impact of popularity bias on items (making popular items even more popular).

[0056] Figure 2 (c) in the diagram represents the causal graph of a recommendation system with popularity bias during the inference phase. Based on the predicted scores obtained during the inference phase, personalized item recommendations can be made to users. Item popularity changes over different historical periods, affecting item attributes, especially for popular items (i.e., items with high popularity), which can easily lead to unstable recommendation results and exacerbate the uneven distribution of item popularity. To address the instability caused by time variations, a method to intervene in item popularity is proposed. A time context reweighting method for item popularity is proposed to control the time period. By intervening in item popularity during the inference phase, a time weight coefficient including time context features is assigned to item popularity, thus realizing the positive effect of time-independent item popularity on recommendation results.

[0057] The training phase of the item recommendation method based on causal popularity debiasing proposed in this invention, i.e., the training process of the recommendation system of this invention, includes the following steps:

[0058] Optimize the user embedding vectors corresponding to users in the training set, the project embedding vectors corresponding to exposed projects, and the project popularity of each project in multiple historical time periods; under the condition of meeting the set training conditions, obtain the optimized user embedding vector set and the optimized project embedding vector set.

[0059] More specifically, after optimization operations, the initial user embedding vector, initial item embedding vector, and item popularity in the training set are updated to optimized user embedding vectors and optimized item embedding vectors. These optimized embedding vectors are then statistically analyzed to form a set.

[0060] The dataset can be evenly divided into multiple historical time periods. The data from the most recent time period T is used for the inference, training, and validation phases of the recommendation model, serving as the test and validation sets. Data from historical time periods before T are used for the training phase, serving as the training set. Each item corresponds to the item popularity of multiple historical time periods, meaning that item popularity is affected by time changes, and each item's popularity is only included in one historical time period.

[0061] As an example, the training set includes at least one user and at least one exposed item. Furthermore, the training set can include user embedding vectors for all users in the recommendation system, item embedding vectors for all items, and corresponding item popularity; it can also include randomly sampled mini-batch user embedding vectors and item embedding vectors from the recommendation system, along with corresponding item popularity.

[0062] In some embodiments of the present invention, the initialized user embedding vector is obtained by vectorizing the user data in the recommendation system based on embedding techniques; the initialized item embedding vector is obtained by vectorizing the exposed item data in the recommendation system based on embedding techniques. The user embedding vector corresponds one-to-one with the user, and the item embedding vector corresponds one-to-one with the item.

[0063] As an example, the historical interaction relationships between users and exposed items (or user embedding vectors and item embedding vectors) in the recommendation system's training set can be displayed using a user-item interaction matrix. Furthermore, the popularity of each item across multiple historical time periods can also be presented in matrix form.

[0064] The use of matrices to represent users, items, user embedding vectors, item embedding vectors, and item popularity in this embodiment of the invention is merely an example and does not constitute a specific limitation.

[0065] As an example, setting training conditions can be setting the number of training sessions or setting the training precision; however, this invention does not impose any specific limitations on these settings.

[0066] Figure 3 This is a flowchart illustrating the optimization operation of the causal popularity-based bias-removal recommendation algorithm of the present invention during the training phase.

[0067] like Figure 3 As shown, the optimization operation includes steps S101-S103:

[0068] Step S101: Select any project, select the project popularity in any historical time period corresponding to the project, select any user, and initialize the project embedding vector corresponding to the project and the user embedding vector corresponding to the user.

[0069] More specifically, selecting any item in the recommendation system can be achieved by initializing the item embedding vector, and then by selecting the item popularity for any historical time period corresponding to that item (for example, selecting the item popularity of item i in historical time period t). This allows for intervention in the project; any user selected from the recommendation system can be represented by an initialized user embedding vector, which facilitates the quantification of user consistency in the future.

[0070] Step S102: Obtain a matching score based on the selected item popularity, user embedding vector, item embedding vector and pre-stored recommendation model, and obtain a consistency score based on the selected item popularity, user embedding vector and pre-stored user consistency model.

[0071] In the process of obtaining the matching score and consistency score in step S102, the user embedding vector and the item embedding vector can be used to represent the user and the item respectively for calculation.

[0072] Figure 4 This diagram illustrates the time-aware causal popularity debiasing framework for recommendation algorithm (TAPD) proposed in this invention. Figure 4 As shown, a matching score can be obtained based on the selected item popularity, user embedding vector, item embedding vector, and a pre-stored recommendation model; a consistency score can be obtained based on the selected item popularity, user embedding vector, and a pre-stored user consistency model. The pre-stored recommendation model can be a traditional recommendation model or a custom recommendation model. Traditional recommendation models are recommendation system algorithms, including collaborative filtering, matrix factorization, logistic regression (LR), factorization machines (FM), field-aware factorization machines (FFM), and gradient boosting decision tree (GBDT), etc.

[0073] In some embodiments of the present invention, a matching score is obtained based on the selected item popularity, user embedding vector, item embedding vector, and a pre-stored recommendation model, including:

[0074] In the recommendation module, the selected user embedding vector and item embedding vector are input into a pre-stored traditional recommendation model. The model output is then multiplied by the popularity of the selected items to obtain a matching score.

[0075] As an example, taking the recommendation model as a matrix factorization (MF) model, the formula for removing confounding factors (item popularity being a confounding factor between the item and the recommendation result) through the do-calculus in causal inference on the item side can be expressed as:

[0076]

[0077] Among them, f ui e represents the matching score between user u and item i. u Let e ​​represent the embedding vector of user u. i This represents the embedding vector of item i. ρ represents the popularity of item i in the historical time period t, and ρ is a hyperparameter used to represent the time influence coefficient to reflect the time-aware characteristics of the recommendation process.

[0078] As an example, if the pre-stored recommendation model is a custom model, item popularity, user embedding vector, and item embedding vector can also be input into the recommendation model, and the model will directly output a matching score. This invention does not impose specific limitations on the calculation method for obtaining the matching score from the item popularity, user embedding vector, and item embedding vector based on the recommendation model.

[0079] In some embodiments of the present invention, a consistency score is obtained based on the selected item popularity, user embedding vectors, and a pre-stored user consistency model, including:

[0080] The selected project popularity and user embedding vectors are input into a pre-stored user consistency model to calculate the value of the user embedding vector divided by the sum of the project popularity and a set constant, and this value is used as the consistency score.

[0081] During the training phase, user consistency is quantified on the user side to represent the correlation between project popularity and users. Referring to the inverse propensity scoring (IPS) method, user consistency can be expressed by the formula:

[0082]

[0083] Among them, c ui β represents the consistency score between user u and project i, and β is a set constant used to avoid the risk of division by zero.

[0084] Step S103: Obtain the training objective function based on the matching score and consistency score, and iteratively optimize the user embedding vector and item embedding vector based on the training objective function.

[0085] More specifically, such as Figure 4 As shown in the recommendation framework, based on the matching score f ui The recommended loss function L can be obtained. I Based on the consistency score c ui The consistency loss function L can be obtained. U According to the consistency loss function L U Recommended loss function L I The training objective function can be obtained by setting the hyperparameter α.

[0086] In some embodiments of the present invention, the training objective function is obtained based on the matching score and the consistency score, including:

[0087] Based on the matching scores obtained from positive and negative samples, a recommendation loss function is constructed using the Bayes Personal Rank (BPR) algorithm; where positive samples are items that have historical interactions with the user, and negative samples are items that have not historical interactions with the user.

[0088] The consistency loss function is derived based on the binary cross-entropy loss function and the consistency score.

[0089] The training objective function is obtained based on the consistency loss function, the recommendation loss function, and the setting of hyperparameters.

[0090] More specifically, the recommendation training phase includes two parts: the user side and the item side. The training objective function not only needs to consider the negative impact of popularity bias that needs to be removed on the item side, i.e. popular items are more likely to have a higher exposure probability due to the increase in item popularity, but also needs to consider the negative impact of user consistency on the recommendation results on the user side.

[0091] As an example, this invention mainly constructs a recommendation loss function for the causal popularity debiasing process based on the Bayesian personalized ranking algorithm. Therefore, the recommendation loss function L on the project side... I The formula can be expressed as:

[0092]

[0093] in, Let f represent the training set. ui f represents the matching score between positive sample i and user u. uj Let represent the matching score between negative sample j and user u, positive sample i refers to items that have historical interaction with user u, negative sample j refers to items that have no historical interaction with user u, σ(·) is the sigmoid function, and n represents the number of users in the recommendation system.

[0094] The above recommended loss function L I f in the formula ui and fuj It must be obtained based on the same user embedding vector corresponding to the same user u.

[0095] The consistency loss function L on the user side is used to mitigate the user consistency goal. U The formula can be expressed as:

[0096]

[0097] The training objective function L during the training phase can be expressed by the following formula:

[0098] L = L I +α·L U ;

[0099] Here, α is a hyperparameter used to balance the user side and the project side; and for the sake of simplicity, the regularization term is omitted here.

[0100] In some embodiments of the present invention, the user embedding vector and item embedding vector selected based on iterative optimization of the training objective function include:

[0101] The Adaptive Moment Estimation (Adam) algorithm, also known as the Adam optimizer, is used to iteratively optimize the feature values ​​in the user embedding vector and the item embedding vector based on the training objective function.

[0102] With the help of do calculus, Figure 2 The derivation of the training process for item (b) in the table can be expressed as follows:

[0103]

[0104] Here, U represents the user set, I represents the item set, and R represents the binary recommendation result (e.g., if the recommendation result is "recommended," then the value of R is 1; if the recommendation result is "not recommended," then the value of R is 0). The first equation is based on the existence of a sufficient set of confounding factor item popularity Z. Using Z as a condition, the Z→I path is blocked. At this point, there is no path between I and R that points to I, satisfying the backdoor criterion, and the backdoor adjustment formula can be used. The second equation is because after intervening in I, the item popularity Z will be independent of I, and Z and U are unaffected by the intervention due to the lack of causal effect. The third equation utilizes the matching score f through transformation. ui Estimated recommended causal probability In the fourth equation, E(Z) is the expectation of Z, which is the same for all interaction results and can be ignored. Ultimately, this conditional probability can be expressed by f. ui Estimate the causal effect of the recommendations.

[0105] Project popularity changes over different historical periods Figure 2 The path from time → item popularity Z → binary recommendation result R in the inference process leads to instability in the recommendation results. Most existing models only consider the impact of a portion of the time period on popularity, neglecting the instability caused by the change in item popularity over time. For example, the PDA framework randomly selects item popularity over a fixed time period, assuming that the item popularity of the latest time period and its adjacent time periods have a significant impact on the recommendation results; the TIDE framework believes that the factors influencing popularity bias are time-sensitive user consistency and time-independent item intrinsic quality, and that the strength of consistency bias is affected by the cumulative estimation of the current time encoding and its previous time encodings. Since item popularity is affected by time changes, this invention believes that the temporal context can reflect more potential information. By intervening in item popularity during the inference stage, the positive effect of popularity bias on the recommendation results can be leveraged, making the recommendation results more accurate.

[0106] Assume that users can only interact with items already displayed in the recommendation system. In some embodiments of the invention, such as... Figure 5 As shown, the specific steps in the inference phase of a recommendation system based on causal popularity bias removal include:

[0107] Step S201: Based on the popularity of each exposed item in the recommendation system across multiple historical time periods, calculate the time context weight coefficient for each historical time period.

[0108] In some embodiments of the present invention, based on the project popularity of each exposed project across multiple historical time periods, a time context weight coefficient corresponding to each historical time period is calculated, including:

[0109] Calculate the mean popularity of all items for each historical time period t. Average popularity of projects in various historical periods The ratio of the project popularity to the sum of the average popularity of projects across all historical time periods is used as the temporal context weighting coefficient for each historical time period; or

[0110] Calculate the difference between the mean popularity of all items in each historical time period t and the mean popularity of all items in the previous historical time period t-1. The difference between the mean popularity of the project in different historical time periods The ratio of the sum of the differences between the project popularity and the average popularity of the project across all historical time periods is used as the temporal context weighting coefficient for each historical time period.

[0111] As an example, if there are T+1 historical time periods, where T represents the most recent time period and 0 represents the oldest time period, then the formula for the time context weight coefficient can be expressed as:

[0112]

[0113] Where T-1 represents the time period preceding the latest time period T.

[0114] `sim__w(t)` directly calculates the average percentage of project popularity within each historical time period up to and including time period T-1. `linear__w(t)` calculates the linear difference in popularity between all time periods.

[0115] Step S202: Multiply the project popularity and time context weight coefficient of each historical time period corresponding to each exposed project, and sum the products to obtain the fixed project popularity corresponding to each project.

[0116] More specifically, the popularity of the project corresponding to project i within the historical time period t. By assigning time context weight coefficients to the corresponding historical time periods and summing the products of each historical time period (i.e. controlling the causal influence of all historical time periods on the popularity of the project), the fixed popularity zi′ of project i can be obtained.

[0117] As an example, when calculating the popularity of a fixed item, the entire recommendation system can only choose sim__w(t) or linear__w(t) as the temporal context weight coefficient for the popularity of all items in the historical time period t.

[0118] Step S203: Based on the pre-stored recommendation model, fixed item popularity, optimized user embedding vector set, and optimized item embedding vector set, obtain the predicted score between the user and the exposed items, and recommend exposed items to the user based on the predicted score; wherein, the optimized user embedding vector set and optimized item embedding vector set are obtained through the training phase.

[0119] In some embodiments of the present invention, a predicted score between a user and exposed items is obtained based on a pre-stored recommendation model, fixed item popularity, an optimized user embedding vector set, and an optimized item embedding vector set, including:

[0120] Select any optimized user embedding vector e′ from the optimized user embedding vector set and the optimized project embedding vector set. u and optimize the project embedding vector e′ i Input it into a pre-stored traditional recommendation model, and compare the model results with the fixed item popularity z corresponding to item i. i Perform a dot product operation to obtain the predicted scores between users and items corresponding to the selected optimized user embedding vector and optimized item embedding vector, respectively;

[0121] Repeat the above steps to obtain the predicted scores between all users and items corresponding to the optimized user embedding vector set and the optimized item embedding vector set, which are the predicted scores between all users and all items in the recommendation system.

[0122] When calculating the predicted score, the pre-stored recommendation model in step S203 of the inference phase is the same model as the recommendation model mentioned in step S102 of the training phase.

[0123] As an example, if users and items are displayed in the form of a user-item interaction matrix, after obtaining optimized user embedding vectors and optimized item embedding vectors during the training phase, the optimized embedding vectors can be used to update the pre-training embedding vectors in the user-item interaction matrix, thereby obtaining the optimized user-item interaction matrix. That is, the optimized user embedding vector set and the optimized item embedding vector set can be displayed in the form of an optimized user-item interaction matrix.

[0124] In the process of obtaining predicted scores during the inference phase, user embedding vectors and item embedding vectors can be used to represent users and items respectively for calculation. For example... Figure 4 As shown, the prediction scores obtained during the inference phase Unaffected by time, taking the MF (Mean-Mi) recommendation model as an example, the predicted score during the recommendation process... The calculation process can be expressed by the formula:

[0125]

[0126] Among them, e′ u This represents optimizing the user embedding vector, e′ i z represents the optimized item embedding vector. i ′ represents the prevalence of the project after intervention, i.e., the prevalence of the fixed project.

[0127] In some embodiments of the present invention, recommending exposed items to users based on predicted scores includes: if the recommendation system includes at least two exposed items, then recommending and sorting the corresponding users according to the predicted scores from largest to smallest, based on the items corresponding to the optimized item embedding vectors in descending order.

[0128] More specifically, the recommendation system can select a set number of items to recommend to the corresponding user u, based on the predicted scores between user u and item i in descending order.

[0129] The reasoning phase was achieved Figure 2 The project popularity intervention process in (c) can be derived using the do calculus as follows:

[0130]

[0131] In this approach, based on interventions to the project during the training phase, the project popularity Z is intervened to make Z independent of time and to block potential confounding factors between Z and R. Time is also independent of U and I, and unaffected by the intervention. The simplified conditional probability can be obtained through... Make an estimate.

[0132] In the causal popularity-based bias removal recommendation algorithm of this invention, the user consistency module is used to remove user consistency. Taking user u and item i as an example, a quantified user consistency recommendation score c is obtained according to the user consistency model. ui Finally, the user consistency loss L is obtained. U The recommendation module adds the popularity of item i over a historical time period t to the traditional recommendation model. Get the matching score f ui By keeping the popularity of items constant and removing the influence of item popularity on items, the recommendation results during the training phase can be represented by the causal probability P(R│U,do(I)), reflecting the intervention process of the recommended items. During the inference phase, the predicted score is calculated using the pre-trained embedding vectors. At this point, the fixed popularity of item i is (z i ′) ρ In this stage, the popularity of the project Z is independent of the historical time period, which removes the influence of time instability and reflects the intervention process of the popularity of the recommended project. The recommendation results in the inference stage can be expressed by causal probability as P(E│U,do(I),do(Z)).

[0133] The item recommendation method and system based on causal popularity debiasing proposed in this invention have the following advantages:

[0134] (1) Time-aware recommendation causal graphs have a wide range of applications, making the recommendation process more interpretable than traditional data-driven analysis methods. They are suitable for real recommendation scenarios and compatible with offline experimental evaluation environments. They intuitively and clearly show the causal relationships between important variables related to the recommendation system, which is helpful for subsequent analysis.

[0135] (2) User consistency modeling is correlated with item popularity. This invention takes into account the impact of item popularity on user interests, combines user embedding vectors with item popularity information to quantify user consistency, alleviate inaccurate user interests caused by users blindly choosing popular items, and adds user consistency loss during the training phase, which is weighed against the recommendation accuracy target, so that the model can learn more realistic and accurate user interests.

[0136] (3) Obtaining time weight coefficients using time context features. This invention proposes a method for representing the time weight coefficients of project popularity from both mean and linear perspectives, reflecting the impact of different time periods on project popularity, and removing the influence of time instability during the inference stage to improve recommendation quality.

[0137] Corresponding to the above method, the present invention also provides a popularity-based item recommendation system, which includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0138] This invention also provides a computer program product having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer program product can be a tangible product, such as random access memory (RAM), memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of product known in the art.

[0139] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0140] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0141] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A project recommendation method based on causal popularity bias removal, characterized in that, The method includes the following steps: Based on the project popularity of each exposed project in various historical time periods, a fixed project popularity is obtained for each project. Based on the pre-stored recommendation model, the fixed item popularity, the optimized user embedding vector set, and the optimized item embedding vector set, a predicted score between the user and the exposed items is obtained, and exposed items are recommended to the user based on the predicted score; The optimized user embedding vector and optimized item embedding vector are obtained based on the user embedding vectors corresponding to users in the training set, the item embedding vectors corresponding to exposed items, and the item popularity of each exposed item across multiple historical time periods, through the following optimization methods: A matching score is obtained based on any item embedding vector, the popularity of the item corresponding to the item embedding vector, any user embedding vector, and a pre-stored recommendation model. A consistency score is obtained based on the item popularity, the user embedding vector, and a pre-stored user consistency model. A training objective function is obtained based on the matching score and the consistency score, and the user embedding vector and the item embedding vector are iteratively optimized based on the training objective function. The process of obtaining a matching score based on any item embedding vector, the popularity of the item corresponding to the item embedding vector, any user embedding vector, and a pre-stored recommendation model includes: inputting any user embedding vector and item embedding vector into the pre-stored recommendation model, and performing a dot product operation between the model output and the item popularity to obtain a matching score. The process of obtaining a consistency score based on the project popularity, the user embedding vector, and the pre-stored user consistency model includes: inputting the project popularity and the user embedding vector into the pre-stored user consistency model, thereby calculating the value of the user embedding vector divided by the sum of the project popularity and a set constant, and using it as the consistency score; The step of obtaining the fixed project popularity for each project based on the project popularity of each exposed project in each historical time period includes: calculating the time context weight coefficient for each historical time period based on the project popularity of each exposed project in multiple historical time periods; multiplying the project popularity of each exposed project in each historical time period with the time context weight coefficient, and summing the products to obtain the fixed project popularity for each project.

2. The method according to claim 1, characterized in that, The user embedding vector is obtained by vectorizing user data based on embedding technology; The project embedding vector is obtained by vectorizing the data of the exposed projects based on embedding technology.

3. The method according to claim 1, characterized in that, The step of obtaining the training objective function based on the matching score and the consistency score includes: Based on the matching scores obtained from positive and negative samples, a recommendation loss function is constructed using a Bayesian personalized ranking algorithm; wherein, the positive samples are items that have historical interactions with the user, and the negative samples are items that have no historical interactions with the user. The consistency loss function is obtained based on the binary cross-entropy loss function and the consistency score; The training objective function is obtained based on the consistency loss function, the recommendation loss function, and the setting of hyperparameters.

4. The method according to claim 1, characterized in that, The iterative optimization of the user embedding vector and the item embedding vector based on the training objective function includes: The adaptive moment estimation algorithm is used to iteratively optimize the feature values ​​in the user embedding vector and the item embedding vector based on the training objective function.

5. The method according to claim 1, characterized in that, The process involves calculating the time context weight coefficient for each historical time period based on the project popularity across multiple historical time periods for each exposed project, including: Calculate the mean popularity of all items across all historical time periods, and use the ratio of the mean popularity of items in each historical time period to the sum of the mean popularity of items across all historical time periods as the temporal context weight coefficient for each historical time period; or Calculate the difference between the mean project popularity of each historical time period and the mean project popularity of all projects in the previous historical time period. Use the ratio of the difference between the mean project popularity of each historical time period and the sum of the differences between the mean project popularity of all historical time periods as the time context weight coefficient of each historical time period.

6. The method according to claim 1, characterized in that, The predicted score between a user and exposed items is obtained from the pre-stored recommendation model, the fixed item popularity, the optimized user embedding vector set, and the optimized item embedding vector set, including: Select any optimized user embedding vector and optimized item embedding vector from the optimized user embedding vector set and the optimized item embedding vector set, input them into the pre-stored recommendation model, and perform a dot product operation between the model result and the corresponding fixed item popularity to obtain the predicted scores between users and items corresponding to the selected optimized user embedding vector and optimized item embedding vector, respectively. Repeat the above steps to obtain the predicted scores between all users and items corresponding to the optimized user embedding vector set and the optimized item embedding vector set; The step of recommending exposed items to users based on the predicted scores includes: sorting the recommended users according to the predicted scores from highest to lowest for the items corresponding to the optimized item embedding vectors.

7. A project recommendation system based on causal popularity bias removal, comprising a processor, a memory, and a computer program / instructions stored in the memory, characterized in that, The processor is used to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 6.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 6.

Images