Archive information retrieval recommendation device based on causal decoupling

Abstract

The embodiment of the invention provides an archive information retrieval recommendation device based on causal decoupling. The method is applied to the technical field of recommendation systems, positive and negative influence factors of popularity are deeply analyzed, and a causal directed acyclic graph is constructed to clarify temporal correlation among different factors, so that the influence of popularity on a recommendation result is adjusted from the source; analyzing and researching potential influence factors of the popularity deviation problem by utilizing a causal directed acyclic graph; specifically, the interaction behavior is mainly driven by the personalized demand of a user and the item popularity, and the dual influence effect of the item popularity is attributed to a time-invariant internal quality factor and a time-varying external interference factor; according to the causal directed acyclic graph, project related attributes such as project popularity feature representation are used as auxiliary information, and influence factors are explicitly separated into hidden variables; a causal decoupling model is constructed, influence factors are used as hidden variables to realize deentanglement, and the influence of popularity on a recommendation result is more accurately adjusted.

Description

Technical Field

[0001] The present disclosure relates to the technical field of recommendation systems, and in particular to an archive information retrieval and recommendation device based on causal decoupling. Background Art

[0002] In the information age, recommender systems have become a key tool for addressing information overload and are widely used in platforms such as e-commerce and online streaming services. However, these systems often suffer from popularity bias. Interaction data often exhibits an unbalanced or long-tail distribution, where a small number of popular items are overexposed to users, while the majority of items are rarely seen. Models trained on this unbalanced data often amplify these biases, forming a feedback loop that leads to problems such as the Matthew effect and echo chambers. As a result, recommendation results become increasingly homogenized, weakening the system's ability to accurately understand users' true preferences and thus degrading the user experience.

[0003] Existing research methods for solving the popularity bias problem can be roughly divided into the following categories: methods based on inverse propensity weighting, methods based on regularization, methods based on reranking, and methods based on causal reasoning. Among them, methods based on inverse propensity weighting are widely used to adjust recommendation results by estimating the propensity weights of items, assigning higher weights to less popular items to alleviate popularity bias. However, they have the problem of large variance, prompting researchers to adopt training methods based on regularization and reranking, including modeling regularization on the training set or adding regularization terms to the loss function, and fine-tuning the recommendation ranking by considering the user's tolerance for favorite items. Methods based on causal reasoning mitigate popularity bias in model training and improve recommendation accuracy by adjusting popularity during the inference stage.

[0004] However, most existing research addresses popularity bias in a static setting, assuming that item popularity and user preferences remain constant over time. This attempt to completely eliminate the impact of popularity bias and evaluate models through a single round of recommendations. However, this setup does not meet the requirements of real-world interactive scenarios. In the real world, the recommendation process is inherently dynamic, with both item popularity and user preferences changing over time. Existing methods for addressing popularity bias lack the processing and analysis of temporal information, hindering the model's ability to timely perceive popularity changes and influencing factors. This inability to effectively mitigate popularity bias in real-world scenarios, resulting in an inability to accurately capture users' true preferences and reduced recommendation accuracy.

[0005] Therefore, in view of the above problems, there is an urgent need to provide an archival information retrieval and recommendation device based on causal decoupling. Summary of the Invention

[0006] The present disclosure provides an archival information retrieval and recommendation device based on causal decoupling, which solves the technical problems in existing technical solutions of completely eliminating the influence of popularity deviation, lacking dynamic perception of popularity, and being unable to accurately capture users' true preferences.

[0007] The present disclosure provides an archival information retrieval and recommendation device based on causal decoupling, comprising: a data acquisition module for constructing a project interaction sequence and a user interaction sequence based on acquired user historical rating time series data and project historical time series data, calculating project popularity feature representation and user satisfaction feature representation based on the project interaction sequence and user interaction sequence, and calculating the Pearson correlation coefficient between the project popularity feature representation and the user satisfaction feature representation; a causal analysis module, connected to the data acquisition module, configured to construct observed variables based on the user interaction sequences, item interaction sequences, and auxiliary information, construct a causal directed acyclic graph based on the observed variables, and analyze the negative impact of latent variables and the relationship between latent variables and observed variables based on the causal directed acyclic graph; wherein the auxiliary information includes a feature representation of item popularity, a feature representation of user satisfaction, and a Pearson correlation coefficient; A decoupling prediction module is connected to the causal analysis module and is used to construct a causal decoupling model based on the relationship between latent variables and observed variables, input the observed variables into the trained causal decoupling model to decouple the negative factors in the latent variables, output recommendation scores for different projects, and select projects to recommend to customers based on the recommendation scores; the causal decoupling model is constructed based on the hidden Markov chain state principle.

[0008] Furthermore, the project interaction sequence includes first interaction data of each project arranged in chronological order; based on the user historical rating time series data and the project historical time series data, the interaction between each project and different users before a preset time is determined to generate first interaction data corresponding to each project; The user interaction sequence includes second interaction data of each user arranged in chronological order; based on the user historical rating time series data and the project historical time series data, the interaction between each user and different projects before a preset time is determined to generate the second interaction data corresponding to each user; The first interaction data and the second interaction data are both 1 or 0, 1 indicates that there is interaction between the user and the project, and 0 indicates that there is no interaction between the project and the user.

[0009] Furthermore, the item popularity feature is obtained by calculating the difference between the item popularity values corresponding to the two end times within a preset time period; the item popularity is obtained by calculating the product of the user coverage and the interaction frequency within the preset time period; the user coverage and the interaction frequency are calculated based on the item interaction sequence and the user interaction sequence; The user satisfaction feature is obtained by calculating the average value of the user's historical rating time series data before a preset time T.

[0010] Furthermore, the latent variables include personalized demand factors, intrinsic quality factors and external interference factors; the negative factors in the latent variables are external interference factors.

[0011] Furthermore, the hidden Markov chain state principle includes: The state of the hidden variable at the current moment only depends on the state of the hidden variable at the previous moment; The item interaction sequence is related to the individual demand factors, intrinsic quality factors and external interference factors in the latent variables; User interaction sequences are only directly related to personalized demand factors; The recommendation score is generated by personalized demand factors and intrinsic quality factors.

[0012] Furthermore, in the decoupling prediction module, a decoupling prediction module is constructed based on the hidden Markov chain state principle, specifically including: A model construction submodule is used to improve the sequential variational autoencoder based on the hidden Markov chain state principle to obtain a causal decoupling model; the causal decoupling model includes a priori network and a posterior network; the priori network and the posterior network each include an encoder and a decoupling unit; the posterior network also includes a decoder; The model training submodule is used to train the model by using the observation variables reconstructed by the causal decoupling model to be close to the input observation variables; and the constructed causal decoupling model is trained using the optimized objective function to obtain a trained causal decoupling model; The recommendation prediction submodule is used to input the observed variables into the trained causal decoupling model. The auxiliary information in the observed variables is processed by the encoder to generate a feature sequence, which is connected with the latent variable at the previous moment and input into the decoupling unit to obtain the prior latent variable; the observed variables are generated into a feature sequence by the encoder, which is connected with the latent variable at the previous moment and input into the decoupling unit. The decoupling unit outputs the posterior latent variable, which is reconstructed by the decoder to output the recommendation score for the item.

[0013] Furthermore, the decoupling unit includes a gated recurrent unit and two independent fully connected layers, and the two fully connected layers are used to output a mean and a logarithmic variance vector respectively.

[0014] Furthermore, the observed variables are input into the trained causal decoupling model to decouple the negative factors in the latent variables, and this is achieved by optimizing the causal decoupling model through an optimized objective function.

[0015] Furthermore, the causal decoupling model adopts an evidence lower bound function as the objective function, and the evidence lower bound function is composed of reconstruction loss and KL divergence.

[0016] The beneficial effects of the present disclosure are: This paper analyzes the interaction matrix data between users and projects, calculates the feature representation reflecting project popularity and user satisfaction, captures its dynamic changes, fits the real recommendation scenario, and solves the problem of dynamic popularity deviation; This paper deeply analyzes the positive and negative impacts of popularity and constructs a causal directed acyclic graph to illustrate the temporal correlation between different factors. This allows us to fundamentally regulate the impact of item popularity on recommendation results, which is more in line with the complexity and dynamics of real-world recommendation scenarios. Using a causal directed acyclic graph (DAG), we analyze and study the potential influencing factors of popularity bias. Specifically, we find that interactive behavior is primarily driven by user personalized needs and item popularity, while the dual influence of item popularity is attributed to time-invariant intrinsic quality factors and time-varying external interference factors. Using DAGs, we use project-related attributes such as item popularity features as auxiliary information to explicitly separate influencing factors into latent variables. A causal decoupling model was constructed, which used influencing factors as latent variables to achieve disentanglement, more accurately adjust the impact of popularity on recommendation results, and improve the performance of the recommendation system and user experience; specifically, by constructing a priori and posterior networks, the model was optimized using the difference between prior and posterior distributions to further separate negative factors in latent variables and improve recommendation accuracy; the decoupling unit was combined with a gated recurrent unit and two independent fully connected layers to output mean and log-variance vectors, respectively, further improving the model's ability to reconstruct time series data.

[0017] By reformulating the objective function to optimize the sequence autoencoder framework, the model effectively mitigates the negative impact of external interference factors, alleviates popularity deviation in dynamic scenarios, and improves the accuracy of recommendations.

[0018] The present invention discloses a causal decoupling sequence recommendation framework to separate influencing factors from a dynamic perspective, optimize recommendation results, and avoid the problem of inaccurate recommendations caused by static adjustments.

[0019] Based on causal reasoning and hidden Markov models, the present invention separates key latent variables that affect user preferences (such as personalized needs and intrinsic quality) from external interference factors, effectively reducing the adverse effects of popularity deviation.

[0020] It should be understood that the contents described in the Summary of the Invention section are not intended to limit the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present disclosure and do not constitute a limitation of the present disclosure. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which: Figure 1 A schematic diagram of the structure of an archive information retrieval and recommendation device based on causal decoupling provided by the present disclosure is shown; Figure 2 A schematic diagram showing the relationship between latent variables and observed variables provided by the present disclosure is shown; Figure 3 A schematic diagram of a decoupled prediction module provided by the present disclosure is shown.

[0022] In the figure, 100 is an archive information retrieval and recommendation device; 110 is a data acquisition module; 120 is a causal analysis module; 130 is a decoupling prediction module; 131 is a model building submodule; 132 is a model training submodule; and 133 is a recommendation prediction submodule. DETAILED DESCRIPTION

[0023] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0024] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0025] The present disclosure provides an archival information retrieval and recommendation device 100 based on causal decoupling, specifically comprising: The data acquisition module 110 is used to construct a project interaction sequence and a user interaction sequence based on the acquired user historical rating time series data and project historical time series data, calculate the project popularity feature representation and the user satisfaction feature representation based on the project interaction sequence and the user interaction sequence, and calculate the Pearson correlation coefficient between the project popularity feature representation and the user satisfaction feature.

[0026] Specifically, in this disclosure, and Represent the user's historical rating time series data and the project's historical time series data, respectively. Indicates the The historical rating time series data of each user, Indicates the Historical rating time series data for each project.

[0027] remember is the interaction data, where For a multi-hot matrix, data is collected in time order and organized into lists of two tuples. and projects In terms of Indicates that at the preset time There was interaction before, which refers to operations such as clicking and browsing; Indicates that there is no interaction during this period. For the convenience of description, use Represents a user At the preset time Previous interaction data with each item, i.e. user interaction sequences; Display items At the preset time Previous interactions with all users, i.e., item interaction sequences.

[0028] In summary, the project interaction sequence U includes the first interaction situation data of each project arranged in chronological order. Based on the user's historical rating time series data and the project's historical time series data, the interaction situation of each project with different users before the preset time T is judged, and the first interaction situation data corresponding to each project is generated; the user interaction sequence I includes the second interaction situation data of each user arranged in chronological order. Based on the user's historical rating time series data and the project's historical time series data, the interaction situation of each user with different projects before the preset time T is judged, and the second interaction situation data corresponding to each user is generated; wherein, the first interaction situation data and the second interaction situation data are both 1 or 0, 1 indicates that there is interaction between the user and the project, and 0 indicates that there is no interaction between the project and the user.

[0029] project The popularity of an item is defined as the user coverage and interaction frequency The product of , where Indicates that at the preset time Front and Project The proportion of users who interact, Indicates the average interaction frequency in the same time period; to evaluate the dynamic changes in popularity, define the project Short-term popularity for: ;in, is the specified interval length, It is a preset time period, and the interval length can be adjusted according to the data density in the actual application scenario; short-term popularity This is the feature representation of project popularity.

[0030] In summary, the project popularity feature is represented by calculating the difference between the project popularity corresponding to the two end times within a preset time period; the project popularity is obtained by calculating the product of user coverage and interaction frequency within a preset time period; user coverage and interaction frequency are calculated based on the project interaction sequence and user interaction sequence.

[0031] User-to-project Satisfaction with the project The average value of the user's historical rating time series data before the preset time T It represents that it changes dynamically over time and constitutes the user satisfaction feature representation. The user satisfaction feature is obtained by calculating the average value of the user's historical rating time series data before the preset time T.

[0032] The Pearson correlation coefficient between the item popularity feature representation and the user satisfaction feature is calculated using the following formula: , in, represents the average popularity of the items, represents the average value of user satisfaction, represents the popularity of the project in the jth stage, It represents the satisfaction of each user in the jth stage, and there are n stages in total.

[0033] The causal analysis module 120 is connected to the data acquisition module 110 and is used to construct observation variables based on the user interaction sequence I, the project interaction sequence U and the auxiliary information, construct a causal directed acyclic graph based on the observation variables, and analyze the negative impact of latent variables and the relationship between latent variables and observation variables based on the causal directed acyclic graph; wherein the auxiliary information includes the feature representation of project popularity , User satisfaction characteristics and Pearson correlation coefficient r.

[0034] This paper selects two real datasets, Douban-Movie and Amazon-CDs, and conducts empirical analysis, that is, analyzing the feature representation of project popularity from the perspective of time series. and user satisfaction characteristics The distribution range of the Pearson correlation coefficient r between the two items above, divided into long-term and short-term situations, leads to the following conclusions about project popularity and user satisfaction: (1) In the long run, the project popularity characteristics represent It is usually positively correlated with its average rating. As the popularity of a project increases, the average rating generally shows an upward trend, indicating a positive correlation between the two. Higher ratings reflect higher user acceptance, which means that popularity is not entirely harmful, and projects that are persistently popular may be favored because of their higher intrinsic quality.

[0035] (2) In the short term, there is an inconsistency between project popularity and user satisfaction. The Pearson correlation coefficient r for some projects is negative, indicating a weak or even negative correlation between project popularity and user satisfaction. In other words, there may be a situation where popularity suddenly increases but the rating decreases. Projects may be affected by external factors in the short term and gain widespread attention, leading to an increase in user conformity. However, this deviates from the true user preferences and reflects the negative impact of popularity.

[0036] Considering that the interaction between users and projects is mainly driven by personalized demand factors and project popularity. Personalized demand factors reflect user preferences and do not have a two-sided impact. Everyone's preferences are different, and it is sufficient to recommend projects that meet the current user needs. However, project popularity has a dual impact and needs to be distinguished through time-related sequence information; the positive impact is attributed to stable intrinsic quality factors, which depends on the long-term popularity of the project and the average score of user satisfaction; the negative impact is attributed to time-varying external interference factors, such as star effects, public opinion promotion, etc., which are represented by the project popularity characteristics. and user satisfaction characteristics The Pearson correlation coefficient r reflects the high or low level of the relationship.

[0037] As can be seen from the above, the influence of project popularity is two-sided. To alleviate the popularity bias problem, it is not necessary to completely remove the influence of popularity, but to reduce its negative impact and use its positive impact to improve the performance of the recommendation system.

[0038] Based on the above analysis, personalized demand factors, intrinsic quality factors and external interference factors are used as latent variables to construct a causal directed acyclic graph for intuitive observation.

[0039] (1) Constructing observation variables , where U represents the user interaction sequence, I represents the item interaction sequence, B represents auxiliary information, and the observed variables are used to infer or estimate the latent variables.

[0040] Auxiliary information B is used to guide the separation of latent variables, including the feature representation of item popularity , User satisfaction characteristics , the Pearson correlation coefficient r between the above two and the basic matching score between users and items. The basic matching score can be obtained by the traditional recommendation system collaborative filtering model based on the item interaction sequence I and the user interaction sequence U. The basic matching score reflects user preferences and is based on the above personalized demand factors.

[0041] (2) Constructing latent variables , x, y, and z represent personalized demand factors, intrinsic quality factors, and external interference factors, respectively. The negative factor in the latent variable h is the external interference factor z. The latent variable h, derived from the previous empirical analysis, supplements and explains the observed data in the causal diagram. By introducing the latent variable h, we can more comprehensively understand the relationships between variables and construct a more accurate causal diagram.

[0042] Existing algorithms or software can be used to construct the causal directed acyclic graph in the present disclosure, and this application does not limit this.

[0043] The causal directed acyclic graph constructed to describe the dynamic decoupling process of the model can be expressed as the following structural equation:

[0044] , Among them, f x (·), f y (·), f z (·), f U (·), f S (·) and f I (·) is the state transfer function of the implicit function, which is constructed based on the principle of hidden Markov chain; x t represents the personalized demand factor at the current time t, y t represents the intrinsic quality factor at the current time t, z t represents the external interference factors at the current time t, U t represents the user interaction sequence reconstructed at the current time t, S t represents the recommendation score at the current time t, I t represents the project interaction sequence reconstructed at the current time t, B t represents the auxiliary information at the current time t, x t-1represents the personalized demand factor at the previous moment t-1, y t-1 represents the intrinsic quality factor at the previous moment t-1, z t-1 Represents the external interference factor at the previous moment t-1.

[0045] According to the constructed causal directed acyclic graph, we can know the following hidden variables h and observed variables Relationship: The state of the hidden variable at the current moment only depends on the state of the hidden variable at the previous moment; The item interaction sequence I is related to the personalized demand factor x, the intrinsic quality factor y and the external interference factor z in the latent variable h; The user interaction sequence U is only directly related to the personalized demand factor x, which reduces the adverse effects of external interference factors; To reduce the adverse effects of external interference factors, the recommendation score S is generated by the personalized demand factor x and the intrinsic quality factor y.

[0046] Specifically, see Figure 2 , the hidden variable h at the current time t t Auxiliary information B t The influence of, and the current hidden variable h t The state of only depends on the hidden variable h at the previous moment t-1 The state of participates in the reconstruction process of the observed variables while being propagated and updated at each time step, further proving the conclusion of the empirical analysis and indicating that the state changes of the latent variable h in this disclosure can be inferred through the Markov chain model.

[0047] The decoupling prediction module 130 is connected to the causal analysis module 120 and is used to construct a causal decoupling model based on the relationship between latent variables and observed variables, input the observed variables into the trained causal decoupling model to decouple the negative factors in the latent variables, output recommendation scores for different projects, and select projects to recommend to customers based on the recommendation scores; the causal decoupling model is constructed based on the hidden Markov chain state principle.

[0048] Specifically, see Figure 3 In the decoupling prediction module 130, a decoupling prediction module is constructed based on the hidden Markov chain state principle, specifically including: The model construction submodule 131 is used to improve the sequential variational autoencoder based on the hidden Markov chain state principle to obtain a causal decoupling model; the causal decoupling model includes a prior network and a posterior network; the prior network and the posterior network both include an encoder and a decoupling unit; the posterior network also includes a decoder; the prior network and the posterior network share the encoder and decoupling unit.

[0049] The model training submodule 132 is used to train the model by using the observation variables reconstructed by the causal decoupling model to be close to the input observation variables; and to train the constructed causal decoupling model using the optimized objective function to obtain a trained causal decoupling model; The recommendation prediction submodule 133 is used to input the observed variables into the trained causal decoupling model. The auxiliary information in the observed variables is processed by the encoder to generate a feature sequence, which is connected with the latent variable at the previous moment and input into the decoupling unit to obtain the prior latent variable; the observed variables are generated into a feature sequence by the encoder, and are connected with the latent variable at the previous moment and input into the decoupling unit. The decoupling unit outputs the posterior latent variable, and the posterior latent variable is reconstructed by the decoder to output the recommendation score for the item.

[0050] Based on the above analysis, the sequential variational autoencoder is improved according to the state analysis principle of hidden Markov chains. Because the intermediate states in the hidden Markov chain model are hidden and cannot be directly observed, they must be inferred through observed variables. In this case, it is necessary to introduce the state relationship of the hidden Markov chain into the sequential variational autoencoder. That is, the encoder encodes the observed variables into a distribution function of the latent space to obtain the hidden variables. The decoder then reconstructs the input data from this latent space. By comparing the differences between the input data before and after reconstruction, the state transition relationship of the hidden variables is inferred, improving the model's ability to reconstruct time series data.

[0051] The decoupling unit includes a gated recurrent unit and two independent fully connected layers, and the two fully connected layers output the mean and logarithmic variance vectors respectively.

[0052] During the training process, reconstructed observation variables are used to approximate the real observation data, ensuring that the generated variables are close to the actual data, so that the model can better learn the underlying data features and make accurate recommendations.

[0053] The observed variables are input into the trained causal decoupling model to decouple the negative factors in the latent variables. This is achieved by optimizing the causal decoupling model through an optimized objective function. The causal decoupling model uses an evidence lower bound function as the objective function, which is composed of reconstruction loss and KL divergence. Specifically, the reconstruction loss measures the similarity between the data generated by the model and the real data. Maximizing the log-likelihood estimate of the reconstructed data makes the generated data as close as possible to the actual observed data. The formula is as follows: ,in, Represents the log-likelihood expectation of the data generated by the model and the real data, given In the case of The expected value of , E means taking the expected value; reflects the model's ability to reconstruct observed data; represents the posterior latent variable, which means that the model infers the distribution of latent variables based on observed variables. represents the generated distribution, that is, the distribution of the observed variables reconstructed by the model.

[0054] KL divergence measures the posterior latent variables learned by the model and prior latent variables Minimizing the KL divergence ensures that the distribution of latent variables is closer to the prior latent variables. The formula is as follows: ;in, is the prior distribution, which represents the prior hypothesis of the latent variable h. Used to measure the difference between c and d.

[0055] The evidence lower bound function can be expressed as , based on the KL divergence definition expansion term, the simplified evidence lower bound function L formula is as follows: ;in, represents the prior distribution of all latent variables h in the system before the preset time T. By optimizing the above objective function, the distribution parameters of each latent variable h are learned and estimated through the decoupling unit, thereby achieving the decoupling of negative factors in the latent variables.

[0056] Specifically, the causal relationship between the observed variable o and the latent variable h before the preset time T is It can be expressed as follows through the joint distribution: ;in, 、 and They are respectively the user interaction sequence U learned in the decoder at the current time t t , Recommendation score S t Interaction Sequence I t The distribution of, where the recommendation score S t Only with x t and y t It is related to the decoupling of external interference factors. Represents the latent variable h based on the previous moment t-1 t-1 and the auxiliary information B at the current time t t The hidden variable h at the current time t is obtained t The prior distribution of .

[0057] Take the observed variable B at the current time t t and the hidden variable h at the previous moment t-1 t-1 As input, for each moment of the latent variable h, the prior distribution The gated recurrent unit is parameterized as , to capture long-term dependencies in sequential data. Specifically, the gated recurrent unit selectively updates information using reset and update gates, then processes the output through two independent fully connected layers to obtain the mean and logarithmic variance vectors of the latent variables, respectively. This is the role of the decoupling unit.

[0058] Due to the complexity of direct calculation of the prior latent variables, the posterior latent variables are used in the formula Make an approximation:

[0059] in, Represents the observed variable at the current time t and the hidden variable h at the previous moment t-1 t-1 The posterior distribution of the latent variable h at the current moment t obtained under the condition of Indicates that the current time t is at x t and y t S is obtained under the premise of t The posterior distribution of Indicates that before the prediction time T, the observed variables The recommendation score obtained under the premise of The posterior distribution of is used as the denominator to normalize the posterior distribution.

[0060] In a Bayesian framework, the posterior distribution usually captures the conditional probability of the latent variable based on the observed variable. Consistent with the prior hypothesis, the posterior network uses the same reparameterization method as the prior network, transforming the observed variables The feature vector is obtained by inputting the encoder and compared with the posterior latent variable h at the previous moment. t-1 Connection, input to the decoupling unit.

[0061] Decoupling unit output reconstructed observation variables and recommendation score S t ,During the training process, the generated reconstructed variables are used to approximate the real ,observed data, ensuring that the generated variables are close to the ,actual data, so that the model can better learn the underlying data ,features and make accurate recommendations.

[0062] This disclosure also provides a specific embodiment to verify feasibility: This paper uses three widely used recommendation system datasets for specific verification, namely KuaiRec, Douban-Movie and Amazon-CDs.

[0063] Among them, the KuaiRec dataset contains millions of detailed interaction records of a short video platform, including user attributes, video features, interaction timestamps, viewing time ratio and other information; for data with a viewing time ratio of more than 5, it is limited to 5 to maintain a rating scale of 0 to 5.

[0064] The Douban-Movie dataset comes from a movie review website and contains movie-related attribute information, interaction timestamps, and user ratings. The data volume reaches millions, and the rating level ranges from 1 to 5.

[0065] Amazon-CDs is a dataset collected from a platform that covers a wide range of information, including user ratings on a scale of 1 to 5, timestamps, and product attributes. Its wide time range makes it ideal for analyzing long-term user behavior and dynamic preferences. See Table 1 for details of the dataset. Table 1 Information of three recommendation system datasets

[0066] In the embodiment of the present disclosure, the evaluation indicators include: Normalized Discounted Cumulative Gain (NDCG@L) is used to represent the position of the target item in the top L recommendation lists and measure the ranking quality of the recommendation results; Recall@L measures the proportion of successfully recommended items among the first L recommended items. L is set to 5, 20, or 50 for different datasets. Each experiment is randomly run five times, and the average of the five experimental results is taken to ensure credibility.

[0067] In the embodiments of the present disclosure, implementation details include: The existence of a rating is regarded as a positive interaction, and the leave-one-out evaluation task widely used in recommendation systems is adopted. That is, for each user's continuous interaction sequence, the most recent interaction is taken as the test data, the second to last interaction is taken as the validation data, and the remaining items are used for training.

[0068] To ensure the reliability of user reviews, 10-core filtering is applied to the KuaiRec and Douban-Movie datasets to ensure that each user has at least 10 interaction records with the items. Considering the high sparsity of the Amazon-CDs dataset, 5-core filtering is applied to it.

[0069] During training, the parameters of the observation variable feature generation process are randomly initialized and iteratively optimized. The Adam optimizer is used to estimate the hyperparameters, and the weight decay is set to 0.0001, the learning rate is 0.001, and the embedding size is set to 64.

[0070] In order to ensure the fairness of the experimental results, the embodiments of the present disclosure set up other models for experimental comparison, which are called baseline models, and different embedding dimensions are set according to the structure of the baseline model. Since the PDA, DICE and STOSA models contain two sets of embeddings, their embedding dimensions are searched within {32, 64}, and other baseline models are searched within {64, 128}. Parameters such as the maximum sequence length, learning rate, L2 regularization weight and dropout rate are adjusted using grid search. In order to prevent the model from overfitting, the number of epochs is limited, and training is stopped when the performance on the validation set drops 10 times in a row. The embodiments of the present disclosure are implemented in PyTorch and are equipped with an Intel i7 CPU and an Nvidia 3090 GPU.

[0071] In the embodiments of the present disclosure, the following comparisons are made: As shown in Table 2, this paper compares its performance with three groups of baseline model recommendation methods. The first group consists of mainstream static recommendation system methods, including BPR-MF and LightGCN; the second group focuses on recommendation methods that use causal reasoning to address popularity bias, such as CausE, DICE, and PDA; and the third group includes advanced sequential recommendation methods, including BERT4Rec, DT4SR, and STOSA.

[0072] Table 2 Overall performance of the disclosed and baseline models on three datasets

[0073] In Table 2, the best method results are bolded, and the second-best method results are underlined. Compared to the baseline model, our proposed method achieves optimal results across all metrics, particularly on the KuaiRec dataset, with relative improvements ranging from 9.86% to 16.78%. Furthermore, our proposed method demonstrates strong adaptability and robustness when working with projects on diverse datasets, effectively addressing the challenges posed by varying degrees of popularity bias.

[0074] In order to further verify the effectiveness of the proposed method, the embodiment of the present disclosure verifies the effectiveness of decoupling based on the Douban-Movie dataset.

[0075] First, the present disclosure is trained normally to obtain the distribution of latent variables after decoupling. Then, six independent classifiers are trained, each using a different combination of latent variables for prediction. The experimental results are shown in Table 3.

[0076] Table 3 Prediction results of different latent variable combinations

[0077] Compared to the causal decoupling model disclosed in this paper, the prediction results that include the external interference factor z show decreases of 8.4% and 6.2% in recall, and 7.5% and 6.7% in NDCG, respectively. This demonstrates the presence of external interference factors in recommended items and indicates that their inclusion negatively impacts recommendation performance. When the model completely eliminates popularity bias, focusing solely on personalized demand factors for recommendations without considering the intrinsic quality of items, recommendation results deteriorate.

[0078] Experiments show that popularity bias is not entirely harmful, as popular items are generally of higher quality, and incorporating the beneficial impact of popularity can improve recommendation performance. Other experimental results are lower than those of the disclosed method, further confirming the effectiveness of the disclosed method.

[0079] According to the embodiments of the present disclosure, the following technical effects are achieved: This paper analyzes the interaction matrix data between users and projects, calculates the feature representation reflecting project popularity and user satisfaction, captures its dynamic changes, fits the real recommendation scenario, and solves the problem of dynamic popularity deviation; This paper deeply analyzes the positive and negative impacts of popularity and constructs a causal directed acyclic graph to illustrate the temporal correlation between different factors. This allows us to fundamentally regulate the impact of popularity on recommendation results, which is more in line with the complexity and dynamics of real-world recommendation scenarios. Using a causal directed acyclic graph (DAG), we analyze and study the potential influencing factors of popularity bias. Specifically, we find that interactive behavior is primarily driven by user personalized needs and item popularity, while the dual influence of item popularity is attributed to time-invariant intrinsic quality factors and time-varying external interference factors. Using DAGs, we use project-related attributes such as item popularity features as auxiliary information to explicitly separate influencing factors into latent variables. A causal decoupling model was constructed, which uses influencing factors as latent variables to achieve disentanglement, more accurately adjust the impact of popularity on recommendation results, and improve the performance of the recommendation system and user experience; By reformulating the objective function to optimize the sequence autoencoder framework, the model effectively mitigates the negative impact of external interference factors, alleviates popularity deviation in dynamic scenarios, and improves the accuracy of recommendations.

[0080] The present invention discloses a causal decoupling sequence recommendation framework to separate influencing factors from a dynamic perspective, optimize recommendation results, and avoid the problem of inaccurate recommendations caused by static adjustments.

[0081] Based on causal reasoning and hidden Markov models, the present invention separates key latent variables that affect user preferences (such as personalized needs and intrinsic quality) from external interference factors, effectively reducing the adverse effects of popularity deviation.

[0082] In the technical solutions disclosed herein, the acquisition, storage, and application of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0083] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0084] The program code for implementing the technical solutions of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0085] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by an instruction execution device or apparatus or for use in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), optical fibers, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0086] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0087] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0088] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0089] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not limited herein.

[0090] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A device for archival information retrieval and recommendation based on causal decoupling, comprising: A data acquisition module is used to construct a project interaction sequence and a user interaction sequence based on the acquired user historical rating time series data and project historical time series data, calculate the project popularity feature representation and the user satisfaction feature representation based on the project interaction sequence and the user interaction sequence, and calculate the Pearson correlation coefficient between the project popularity feature representation and the user satisfaction feature representation; a causal analysis module, connected to the data acquisition module, configured to construct observed variables based on the user interaction sequences, item interaction sequences, and auxiliary information, construct a causal directed acyclic graph based on the observed variables, and analyze the negative impact of latent variables and the relationship between latent variables and observed variables based on the causal directed acyclic graph; wherein the auxiliary information includes a feature representation of item popularity, a feature representation of user satisfaction, and a Pearson correlation coefficient; A decoupling prediction module is connected to the causal analysis module and is used to construct a causal decoupling model based on the relationship between latent variables and observed variables, input the observed variables into the trained causal decoupling model to decouple the negative factors in the latent variables, output recommendation scores for different projects, and select projects to recommend to customers based on the recommendation scores; the causal decoupling model is constructed based on the hidden Markov chain state principle.

2. The archival information retrieval and recommendation device according to claim 1, wherein: The project interaction sequence includes first interaction data of each project arranged in chronological order; based on the user historical rating time series data and the project historical time series data, the interaction between each project and different users before a preset time is determined to generate first interaction data corresponding to each project; The user interaction sequence includes second interaction data of each user arranged in chronological order; based on the user historical rating time series data and the project historical time series data, the interaction between each user and different projects before a preset time is determined to generate the second interaction data corresponding to each user; The first interaction data and the second interaction data are both 1 or 0, 1 indicates that there is interaction between the user and the project, and 0 indicates that there is no interaction between the project and the user.

3. The archival information retrieval and recommendation device according to claim 1, wherein: The item popularity feature is obtained by calculating the difference between the item popularity at two ends of a preset time period; the item popularity is obtained by calculating the product of user coverage and interaction frequency within the preset time period; The user coverage and interaction frequency are calculated based on the project interaction sequence and the user interaction sequence; The user satisfaction feature is obtained by calculating the average value of the user's historical rating time series data before a preset time T.

4. The archival information retrieval and recommendation device according to claim 1, wherein: The latent variables include personalized demand factors, intrinsic quality factors and external interference factors; the negative factors in the latent variables are external interference factors.

5. The archival information retrieval and recommendation device according to claim 4, wherein: The hidden Markov chain state principle includes: The state of the hidden variable at the current moment only depends on the state of the hidden variable at the previous moment; The item interaction sequence is related to the individual demand factors, intrinsic quality factors and external interference factors in the latent variables; User interaction sequences are only directly related to personalized demand factors; The recommendation score is generated by personalized demand factors and intrinsic quality factors.

6. The archival information retrieval and recommendation device according to claim 5, wherein: In the decoupling prediction module, a decoupling prediction module is constructed based on the hidden Markov chain state principle, which specifically includes: A model construction submodule is used to improve the sequential variational autoencoder based on the hidden Markov chain state principle to obtain a causal decoupling model; the causal decoupling model includes a priori network and a posterior network; the priori network and the posterior network each include an encoder and a decoupling unit; the posterior network also includes a decoder; The model training submodule is used to train the model by using the observation variables reconstructed by the causal decoupling model to be close to the input observation variables; and the constructed causal decoupling model is trained using the optimized objective function to obtain a trained causal decoupling model; The recommendation prediction submodule is used to input the observed variables into the trained causal decoupling model. The auxiliary information in the observed variables is processed by the encoder to generate a feature sequence, which is connected with the latent variable at the previous moment and input into the decoupling unit to obtain the prior latent variable; the observed variables are generated into a feature sequence by the encoder, which is connected with the latent variable at the previous moment and input into the decoupling unit. The decoupling unit outputs the posterior latent variable, which is reconstructed by the decoder to output the recommendation score for the item.

7. The archival information retrieval and recommendation device according to claim 6, wherein: The decoupling unit includes a gated recurrent unit and two independent fully connected layers, and the two fully connected layers are used to output a mean and a logarithmic variance vector respectively.

8. The archival information retrieval and recommendation device according to claim 6, wherein: The observed variables are input into the trained causal decoupling model to decouple the negative factors in the latent variables, and the causal decoupling model is optimized by optimizing the objective function.

9. The archival information retrieval and recommendation device according to claim 8, wherein: The causal decoupling model adopts an evidence lower bound function as the objective function, and the evidence lower bound function is composed of reconstruction loss and KL divergence.

Images