An information recommendation method and device embedded with a causal structure

Abstract

The application discloses an information recommendation method and device embedded with a cause-effect structure, wherein the method comprises the following steps: based on a distribution offset inducement, randomly performing negative sampling on observation data as negative inclination initialization to obtain training data; reconstructing user preference / like / dislike commodity features through a cause-effect structure between user features and user-commodity, exploring a stable cause-effect relationship between the user features, user inclination and negative inclination to learn a data generation mechanism; optimizing fitting of user purchase behavior; optimizing negative sampling probability according to the stable negative inclination; cyclically optimizing in combination until a target function value converges to obtain a stable cause-effect relationship between the user features, user inclination and negative inclination. The application explores a stable cause-effect structure between user features and purchase behavior from observation data, which is not affected by distribution offset, and applies the cause-effect structure to learning stable representation, so that the cause-effect structure can cope with unknown distribution offset in an unknown test environment.

Description

Technical Field

[0001] The present invention relates to the technical fields of recommendation system algorithms, causal structure learning and stable learning theory, and in particular to an information recommendation method and device embedded with a causal structure. Background Art

[0002] In online services such as e-commerce and social media, recommendation systems play a key role in alleviating the growing problem of information overload on websites. The integration of powerful learning models such as deep neural networks and graph convolutional neural networks into recommendation system frameworks has further advanced their development. While current recommendation systems have demonstrated promising performance, most rely on the assumption that training and test data are independent and identically distributed. Unfortunately, due to the unavoidable distribution shift problem in the real world, this assumption is often difficult to meet in real-world recommendation scenarios, significantly impacting the generalization performance of recommendation algorithms.

[0003] The new research results mainly address the challenges brought about by sample selectivity bias, distribution shift, and high sparsity when training recommendation models through data-driven methods, and propose targeted measures in order to learn users' true preferences from the data and the stable causal relationship between them and purchasing behavior, and achieve more stable performance in a test environment with changing data distribution. Summary of the Invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the present invention proposes an information recommendation method embedded with a causal structure, aiming to ensure that the recommendation algorithm maintains stable performance in a test environment where the independent and identically distributed assumption does not hold and there is a shift in the data distribution. It can learn stable user anti-preferences from observed data and restore the generation mechanism between user characteristics and user purchasing behavior, so that the system can recommend products to users more accurately.

[0006] The second object of the present invention is to propose an information recommendation device embedded with a causal structure.

[0007] To achieve the above objectives, an embodiment of the present invention provides an information recommendation method embedded with a causal structure, including:

[0008] S1, based on the inducement of distribution shift, random negative sampling is performed from the observed data as negative tendency initialization to obtain training data;

[0009] S2, based on the training data, reconstructing the product features that users like / dislike through the causal structure between user features and user-product relationships, and discovering the stable causal relationship between user features and user tendencies and negative tendencies to learn the data generation mechanism;

[0010] S3, based on the data generation mechanism, optimizing the fitting of user purchasing behavior;

[0011] S4, obtaining a stable negative tendency of the user based on the fitting of the user's purchase behavior and the learning in S2, and optimizing the negative sampling probability according to the stable negative tendency;

[0012] S5, cyclically jointly optimizing S2 to S4 until the objective function value converges, so as to obtain a stable causal relationship between the user characteristics and the user tendencies and negative tendencies.

[0013] The information recommendation method based on embedded causal structures in the embodiments of the present invention improves the stability of the recommendation algorithm by discovering a stable, distribution-unaffected causal structure between user characteristics and their purchasing behavior from observed data and applying it to learn stable representations, thereby enabling it to cope with unknowable distribution shifts in unknown test environments. This ensures that the recommendation algorithm maintains stable performance in test environments where the independent and identically distributed assumptions do not hold and data distributions are shifted. Stable user anti-preferences are learned from observed data, restoring the generation mechanism between user characteristics and user purchasing behavior.

[0014] In addition, the information recommendation method embedded with causal structure according to the above embodiment of the present invention may also have the following additional technical features:

[0015] The S3 includes: utilizing the stable causal relationship in the data generation mechanism to obtain a stable representation of the user, embedding the stable user representation and the product features into a latent space, optimizing the distance between the user and the product latent representation, and modeling the purchase behavior between the user and the product.

[0016] Furthermore, in one embodiment of the present invention, the method also includes: iteratively alternatingly optimizing an invariant preference learning module based on a causal structure and an anti-preference negative sampling module for processing implicit feedback to learn a stable causal structure between user characteristics and user purchasing behavior and the user's own invariant negative tendency.

[0017] Furthermore, in one embodiment of the present invention, the causes of the distribution shift include: the distribution shift of human behavior caused by the heterogeneity of human behavior at the temporal and spatial levels; and the human intervention of the recommendation system on user behavior.

[0018] Furthermore, in one embodiment of the present invention, the step S4 includes: increasing the probability of extracting products that the user is truly not interested in during negative sampling based on the stable negative tendency.

[0019] To achieve the above objectives, a second embodiment of the present invention provides an information recommendation device embedded with a causal structure, comprising:

[0020] A training module is used to perform random negative sampling from the observed data as negative tendency initialization based on the inducement of distribution shift to obtain training data;

[0021] A first learning module is configured to reconstruct the characteristics of user-preferred / disliked products based on the training data through the causal structure between user characteristics and user-product relationships, and to discover the stable causal relationship between user characteristics and user tendencies and negative tendencies, so as to learn the data generation mechanism;

[0022] A first optimization module, configured to optimize the fitting of user purchasing behavior based on a generation mechanism of the data;

[0023] A second optimization module is configured to obtain a stable negative tendency of the user based on the fitting of the user's purchasing behavior and the learning of the first learning module, so as to optimize the negative sampling probability according to the stable negative tendency;

[0024] The loop module is used to cyclically jointly optimize the first learning module to the second optimization module until the objective function value converges, so as to obtain a stable causal relationship between the user characteristics and the user tendencies and negative tendencies.

[0025] The information recommendation device embedded with a causal structure in the embodiments of the present invention improves the stability of the recommendation algorithm by discovering a stable, distribution-unaffected causal structure between user characteristics and purchasing behavior from observed data and applying it to learn stable representations, thereby enabling it to cope with unknowable distribution shifts in unknown test environments. This ensures that the recommendation algorithm maintains stable performance in test environments where the independent and identically distributed assumptions (IIDs) do not hold and data distributions are shifted. Stable user anti-preferences are learned from observed data, restoring the generation mechanism between user characteristics and user purchasing behavior.

[0026] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0028] Figure 1Flowchart of a method for recommending information embedded with a causal structure according to one embodiment of the present invention;

[0029] Figure 2 A schematic diagram of a collaborative causal filtering algorithm framework according to an embodiment of the present invention;

[0030] Figure 3 Schematic diagram of the structure of an information recommendation device embedded with a causal structure according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0032] The following describes the information recommendation method and device embedded with causal structure according to an embodiment of the present invention with reference to the accompanying drawings. First, the information recommendation method embedded with causal structure according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0033] Figure 1 Flowchart of a method for recommending information embedded with a causal structure according to one embodiment of the present invention.

[0034] like Figure 1 As shown, the method includes but is not limited to the following steps:

[0035] S1, based on the inducement of distribution shift, random negative sampling is performed from the observed data as negative tendency initialization to obtain training data;

[0036] S2, based on the training data, reconstructs the product features that users like / dislike through the causal structure between user features and user-product relationships, and discovers the stable causal relationship between user features and user tendencies and negative tendencies to learn the data generation mechanism;

[0037] S3, based on the data generation mechanism, optimizing the fitting of user purchasing behavior;

[0038] S4, obtaining a stable negative tendency of the user based on the fitting of the user's purchase behavior and the learning in S2, and optimizing the negative sampling probability according to the stable negative tendency;

[0039] S5, cyclically jointly optimize S2 to S4 until the objective function value converges to obtain a stable causal relationship between stable user features and user tendencies and negative tendencies.

[0040] As an example, Figure 2This is a schematic diagram of the collaborative causal filtering algorithm framework according to one embodiment of the present invention. This invention mines stable causal structures from data and optimizes representation learning based on these structures, naturally integrating with mainstream recommendation algorithm frameworks. This ensures that the recommendation algorithm maintains stable performance in test environments where the independent and identically distributed assumptions (IIDs) violate and data distributions are skewed. It also learns stable user anti-preferences from observed data and restores the mechanism that drives the relationship between user characteristics and purchasing behavior.

[0041] The embodiments of the present invention are further described below.

[0042] Understandably, there are two main causes of distribution shift in the real world. First, human behavior exhibits significant temporal and spatial heterogeneity, leading to naturally occurring distribution shifts in human behavior. For example, the distribution of user behavior in coastal cities is very likely different from that in inland cities; and best-selling products in winter are very likely different from those in summer. Another cause is the artificial interference of recommendation systems in user behavior. For example, users tend to pay more attention to popular products or products that are higher on the recommendation list.

[0043] To this end, the specific method of the present invention is:

[0044] 1. First, random negative sampling is performed from the observed data as negative tendency initialization to obtain training data.

[0045] 2. Learning data generation mechanism: Reconstruct the product features that users like / dislike through the causal structure between user features and user-product relationships, thereby discovering the stable causal relationship between user features and their propensity and negative propensity.

[0046] 3. Optimize the fit of user purchasing behavior: Leverage the stable causal relationship in the data generation mechanism to obtain a stable user representation, embed the stable user representation and product features into the latent space, and optimize the distance between the user and product latent representations, that is, model the interaction (purchasing) behavior between users and products.

[0047] 4. Optimize negative sampling based on user negative propensity: After the algorithm in step 2 has learned the user's stable negative propensity, based on this negative propensity, the probability of extracting products that the user is truly not interested in during negative sampling is increased.

[0048] 5. Finally, steps 2 to 4 are jointly optimized cyclically until the objective function value converges, and a stable causal relationship between user features and their tendencies and negative tendencies can be obtained.

[0049] As an example, the present invention investigates and analyzes the instability of the correlation between user preferences and user behaviors learned by a recommendation algorithm in a real scenario with distribution shift.

[0050] As an example, the present invention mines stable causal structures from data and optimizes representation learning based on them, thereby providing a stable recommendation learning framework that improves the robustness of the recommendation algorithm.

[0051] As an example, the present invention iteratively and alternately optimizes an invariant preference learning module based on causal structure and an anti-preference negative sampling module for processing implicit feedback to learn a stable causal structure between user characteristics and their purchasing behavior and the user's own invariant negative tendency.

[0052] The recommendation method based on the above-mentioned method solves the problem that in the field of traditional network services, traditional data-driven recommendation algorithms are highly dependent on the independent and identically distributed assumptions generated by data, while the data sources in real scenarios are unknowable, and in most cases, only positive signals are available, while negative signals are difficult to obtain. In real life, when it is determined that a user likes an item, the user will pay more attention to popular items, and the user will also pay more attention to items at the top of the recommendation list. The system recommends products based on the user's personal preferences without considering the influence of other factors, resulting in the recommendation system recommending products to users with deviations and inaccuracies. The present invention, by discovering a stable causal structure between user characteristics and their purchasing behavior that is not affected by distribution shifts from observed data, and applying it to learning stable representations, can cope with unknown distribution shifts in unknown test environments, making the products recommended to users more accurate.

[0053] In summary, the present invention mines stable causal structures from data and optimizes representation learning based on this, which can be naturally combined with the mainstream recommendation algorithm framework. It iteratively alternately optimizes the invariant preference learning module based on the causal structure and the anti-preference negative sampling module that processes implicit feedback to learn a stable causal structure and the user's invariant negative tendency, which has outstanding advantages in new samples and new environments. It has inherent interpretability by discovering the causal structure in user purchasing behavior. The present invention discovers a stable causal structure between user characteristics and their purchasing behavior that is not affected by distribution shifts from the observed data, and applies it to learning stable representations, so that it can cope with unknown distribution shifts in unknown test environments.

[0054] Figure 3 FIG. 1 is a schematic diagram of a structure of an information recommendation device embedded with a causal structure according to an embodiment of the present invention. Figure 3 As shown, the information recommendation device 10 embedded with a causal structure includes: a training module 100, a first learning module 200, a first optimization module 300, a second optimization module 400 and a loop module 500.

[0055] A training module 100 is configured to perform random negative sampling from observed data as negative tendency initialization based on the inducement of distribution shift to obtain training data;

[0056] The first learning module 200 is used to reconstruct the product features of user preferences / dislikes based on the training data through the causal structure between user features and user-product relationships, and to discover the stable causal relationship between user features and user tendencies and negative tendencies, so as to learn the data generation mechanism;

[0057] A first optimization module 300 is used to optimize the fitting of user purchasing behavior based on the data generation mechanism;

[0058] The second optimization module 400 is configured to optimize the negative sampling probability based on the stable negative tendency of the user obtained by fitting the user's purchase behavior and learning from the first learning module 200;

[0059] The loop module 500 is used to cyclically jointly optimize the first learning module 200 to the second optimization module 400 until the objective function value converges, so as to obtain a stable causal relationship between the user characteristics and the user tendencies and negative tendencies.

[0060] As an example, the first optimization module 300 is further configured to:

[0061] The stable causal relationship in the data generation mechanism is used to obtain a stable representation of the user, and the stable user representation and product features are embedded into the latent space. The distance between the latent representation of the user and the product is optimized to model the purchasing behavior between the user and the product.

[0062] As an example, the apparatus 10 further includes:

[0063] The second learning module is used to iteratively and alternately optimize the invariant preference learning module based on causal structure and the anti-preference negative sampling module that processes implicit feedback to learn the stable causal structure between user characteristics and user purchasing behavior as well as the user's own invariant negative tendency.

[0064] As an example, the causes of the above-mentioned training-based distribution shift include: the distribution shift of human behavior caused by the heterogeneity of human behavior at the temporal and spatial levels; and the human intervention of the recommendation system in user behavior.

[0065] As an example, the second optimization module 400 is further configured to: improve the probability of extracting products that the user is truly not interested in during negative sampling based on the stable negative tendency.

[0066] According to an embodiment of the present invention, an information recommendation device embedded with a causal structure improves the stability of the recommendation algorithm by discovering a stable, distribution-unaffected causal structure between user characteristics and their purchasing behavior from observed data and applying it to learn stable representations, thereby enabling it to cope with unknowable distribution shifts in unknown test environments. This ensures that the recommendation algorithm maintains stable performance in test environments where the independent and identically distributed assumptions (IIDs) do not hold and data distributions are shifted. Stable user anti-preferences are learned from observed data, restoring the generation mechanism between user characteristics and user purchasing behavior.

[0067] It should be noted that the aforementioned explanation of the information recommendation method embedded in a causal structure is also applicable to the information recommendation device embedded in a causal structure of this embodiment, and will not be repeated here.

[0068] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0069] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0070] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. An information recommendation method embedded in a causal structure, characterized in that: The following steps are involved: S1, based on the inducement of distribution shift, random negative sampling is performed from the observed data as negative tendency initialization to obtain training data; S2, based on the training data, reconstructing the product features that users like / dislike through the causal structure between user features and user-product relationships, and discovering the stable causal relationship between user features and user tendencies and negative tendencies to learn the data generation mechanism; S3, based on the data generation mechanism, optimizing the fitting of user purchasing behavior; S4, obtaining a stable negative tendency of the user based on the fitting of the user's purchase behavior and the learning in S2, and optimizing the negative sampling probability according to the stable negative tendency; S5, cyclically jointly optimizing S2 to S4 until the objective function value converges, so as to obtain a stable causal relationship between the user characteristics and the user tendencies and negative tendencies.

2. The information recommendation method embedded with causal structure according to claim 1, characterized in that: The S3 includes: The stable causal relationship in the data generation mechanism is used to obtain a stable representation of the user, and the stable user representation and the product characteristics are embedded in the latent space, and the distance between the user and product latent representations is optimized to model the purchasing behavior between the user and the product.

3. The information recommendation method embedded with causal structure according to claim 1, characterized in that: The method further comprises: The invariant preference learning module based on causal structure and the anti-preference negative sampling module for processing implicit feedback are iteratively optimized to learn the stable causal structure between user characteristics and user purchasing behavior as well as the user's own invariant negative tendency.

4. The information recommendation method embedded with causal structure according to claim 1, characterized in that: The causes of the distribution shift include: Based on the heterogeneity of human behavior at the temporal and spatial levels, the distribution shift of human behavior is formed; and the human intervention of the recommendation system on user behavior.

5. The information recommendation method embedded with causal structure according to claim 1, characterized in that: Said S4 comprises: Based on the stable negative tendency, the probability of extracting products that the user is truly not interested in during negative sampling is improved.

6. An information recommendation device embedded with a causal structure, characterized in that: include: A training module is used to perform random negative sampling from the observed data as negative tendency initialization based on the inducement of distribution shift to obtain training data; A first learning module is configured to reconstruct the characteristics of user-preferred / disliked products based on the training data through the causal structure between user characteristics and user-product relationships, and to discover the stable causal relationship between user characteristics and user tendencies and negative tendencies, so as to learn the data generation mechanism; A first optimization module, configured to optimize the fitting of user purchasing behavior based on a generation mechanism of the data; A second optimization module is configured to obtain a stable negative tendency of the user based on the fitting of the user's purchasing behavior and the learning of the first learning module, so as to optimize the negative sampling probability according to the stable negative tendency; The loop module is used to cyclically jointly optimize the first learning module to the second optimization module until the objective function value converges, so as to obtain a stable causal relationship between the user characteristics and the user tendencies and negative tendencies.

7. The information recommendation device embedded with causal structure according to claim 6, characterized in that: The first optimization module is further configured to: The stable causal relationship in the data generation mechanism is used to obtain a stable representation of the user, and the stable user representation and the product characteristics are embedded in the latent space, and the distance between the user and product latent representations is optimized to model the purchasing behavior between the user and the product.

8. The information recommendation device embedded with causal structure according to claim 6, characterized in that: The device further comprises: The second learning module is used to iteratively and alternately optimize the invariant preference learning module based on causal structure and the anti-preference negative sampling module that processes implicit feedback to learn the stable causal structure between user characteristics and user purchasing behavior as well as the user's own invariant negative tendency.

9. The information recommendation device embedded with causal structure according to claim 6, characterized in that: The causes of distribution shift include: Based on the heterogeneity of human behavior at the temporal and spatial levels, the distribution shift of human behavior is formed; and the human intervention of the recommendation system on user behavior.

10. The information recommendation device embedded with causal structure according to claim 6, characterized in that: The second optimization module is further configured to: Based on the stable negative tendency, the probability of extracting products that the user is truly not interested in during negative sampling is improved.

Images