A recommendation algorithm based on causal inference to alleviate the echo chamber problem

Abstract

This invention relates to a recommendation algorithm based on causal inference to mitigate the echo chamber problem. It eliminates spurious correlations between user representations, item representations, and model predictions, thus eliminating erroneous causal relationships caused by bias. First, causal graph analysis reveals that spurious correlations are caused by item exposure bias due to historical imbalances. Second, a backdoor adjustment method using counterfactual inference is employed to correct bias and avoid the impact of exposure bias. This invention, from a causal perspective, clarifies the mechanisms of interaction and outcome variation among variables, uncovers the true causes of the echo chamber problem, removes confounding factors between user representations, item representations, and predictions, and further reduces the impact of the echo chamber problem on the recommendation system.

Description

Technical Field

[0001] This invention belongs to the field of recommendation system technology, specifically relating to a recommendation algorithm based on causal inference to alleviate the echo chamber problem. Background Technology

[0002] With the explosive growth of internet content, users face the problem of information overload. To improve user experience, personalized recommendation systems have emerged. These systems use users' historical behavioral data (such as browsing history, clicks, searches, purchases, etc.) and behavior on social networks to predict and infer users' interests and preferences, providing customized content and services.

[0003] However, while this data-driven recommendation system based on users' historical behavior increases user engagement, it also presents a potential problem: the system increasingly tends to recommend content that aligns with users' existing interests, thus reducing opportunities for users to access new or heterogeneous information, trapping them in an echo chamber. The echo chamber problem leads people in a homogenized information environment to increasingly seek out content consistent with their existing viewpoints, beliefs, or interests, resulting in reinforced opinions and a weakening of information diversity and heterogeneity.

[0004] Current methods for mitigating the echo chamber problem are based on the idea in traditional machine learning to establish relationships between variables in a recommender system using historical user-item interaction data. The model is trained as a black box structure; simply inputting user information yields recommendations that match their interests and preferences. These methods use neural networks to model the data relationships between users and items, focusing on how to better fit the observed data.

[0005] While this approach demonstrates excellent predictive accuracy, it fails to elucidate the mechanisms underlying the interactions between variables and the resulting variations, making it difficult to explain the effectiveness of existing methods in mitigating the echo chamber problem. Furthermore, considering the echo chamber problem from the perspective of correlations makes it susceptible to spurious correlations, and confounding factors within the system hinder further mitigation of the echo chamber problem from impacting recommendation performance. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a recommendation algorithm based on causal inference to alleviate the echo chamber problem. This invention clarifies how variables in the recommendation system interact under the influence of the echo chamber and the mechanism that leads to changes in the recommendation results, thereby designing a more interpretable and effective recommendation scheme to alleviate the echo chamber problem.

[0007] The technical problem solved by this invention is achieved through the following technical solution:

[0008] A recommendation algorithm based on causal inference to alleviate the echo chamber problem is proposed. The algorithm eliminates spurious correlations between user representations, item representations, and model prediction results, and eliminates erroneous causal relationships caused by bias. First, through causal graph analysis, it is found that the spurious correlations are caused by item exposure bias due to historical imbalance distribution. Second, a backdoor adjustment method using counterfactual reasoning is adopted to avoid the impact of exposure bias.

[0009] Furthermore, by analyzing the underlying mechanism of the echo chamber problem through causal graph analysis, the ideal prediction result Y should only be directly affected by the user representation U and the item representation I, that is, it should only include the paths U→Y and I→Y. These two paths match the user characteristics and item characteristics, and the result is derived from the user's true preferences. However, in reality, there is not only a path U→Y from U to Y, but also a path U←S→Y. This indicates that the user's choice is affected by bias and will be more inclined to interact with items that have already expressed preferences.

[0010] Similarly, the path from I to Y not only exists as I→Y, but also as I←E→Y. This indicates that projects are affected by bias, and projects with higher exposure will be further shown to users. The combination of these two factors results in the observed user interactions consisting of more already interacted projects, i.e., falling into an echo chamber. A qualified model should avoid amplifying bias and rely on genuine user preferences for recommendations, rather than being influenced by bias.

[0011] Furthermore, this paper analyzes how to represent the interaction probability given a user u, item i, user selection tendency s, and item exposure e in a traditional recommendation model. The task is described using a formal language, denoted as P(Y|u,i), and the formula is transformed as follows:

[0012]

[0013]

[0014] A formal description of the biased model P(Y|do(u,i)) after intervention is given, and based on the relevant theory of causal inference, the following can be directly derived:

[0015]

[0016] Furthermore, by employing a backdoor adjustment and bias removal method using counterfactual reasoning, for user embeddings in the model, the user ID representation and user selection representation affected by selection bias are removed, retaining the remaining user feature representations. This achieves the goal of removing selection bias while utilizing user features. In the user embedding, the green circle represents the user ID embedding p. id The yellow circle represents the user-selected embedded representation p. cThe red circle represents the user's remaining feature embedding p. f ;

[0017] First, in the counterfactual world—a world where users dislike items they have expressed preference for—we obtain counterfactual user embeddings that do not include selection embeddings, reducing the impact of selection bias. Since user ID embeddings have a significant impact on recommendation results and are also affected by selection bias, we need to set the degree to which user ID embeddings are removed, represented by dashed circles. For item embeddings in the model, a backdoor adjustment method is used, performing causal inference with the aid of a causal graph. Because traditional recommendation models do not quantify the impact of exposure probability on recommendation results, we calculate the exposure probability of each item and then remove the exposure embedding from the item representation using the quantified exposure probability. White circles represent item ID embeddings q. id The blue circle represents the feature embedding q of the project. f The black circle indicates the project's exposure embedding. This represents the exposure of item i at time t. Then, a neural network is trained, and the result after training is the recommendation result after removing user selection bias and item exposure bias.

[0018] Then, counterfactual reasoning is used to remove the spurious correlation between user representations and prediction results. Users tend to interact with content they like, so most of the user's historical interaction records are content that the user is interested in or likes. Content that they dislike or are not interested in is often not recorded. This leads to the model overlearning user preference items, inducing users to click more on content similar to their own preferences. This results in selection bias, that is, users are more inclined to interact with items that are similar to their expressed preferences. By using counterfactual thinking, "If the user has not expressed a liking for this type of item, then what items should the recommendation system recommend to the user?" Based on this idea, the system helps users step out of the interest circle of their user group with similar interests, i.e., the echo chamber, reducing the recommendation of content that the user group is interested in, avoiding the user being induced by selection, and making them interact with items based on their own true preferences, thus avoiding the impact of selection bias.

[0019] We use counterfactual reasoning to imagine "what would the prediction result be if the user representation did not include a user ID?" We describe this using formal language, and the prediction result obtained in the real world is represented as Y. u,i Then the prediction obtained in the counterfactual world is: in The user representation does not include the user ID representation. To obtain the effect of the user ID representation, simply use... Therefore, a weighting factor α is used to weigh the prediction results after subtracting the influence of the user ID:

[0020]

[0021] Finally, a backdoor adjustment strategy is employed to eliminate the spurious correlation between project representations and prediction results, and the parameters are learned based on the user's interaction history and the BPR loss function used in implicit feedback.

[0022]

[0023] To express P θ 9y=1|u,i,e) can be obtained by simultaneously inputting the user representation $U$, the project representation $I$, and the project exposure representation into the new model. However, considering that integrating user-project matching and exposure together would require reconsidering the integration scheme once the model is changed, it lacks scalability. Therefore, we consider decoupling user-project matching from project exposure:

[0024]

[0025] Considering P θ (y|u,i,e) is a positive value, therefore the model result f needs to be adjusted. θ (u,i) remains positive, therefore a variant of the ELU activation function, ELU', is used, which is specifically expressed as follows:

[0026]

[0027] Regarding the project's exposure Divide user interaction history with the project into time segments and record... Record user interactions with the project at time t. This represents the record of item i being interacted with at time t. The exposure of item i at time t is thus represented as:

[0028]

[0029] in: This represents the number of times item i is observed at time t. It is obtained by summing all user interactions with item i during time t. This represents the sum of the number of interactions with all items within time t;

[0030] This leads to the final training loss function that removes project exposure bias:

[0031]

[0032] Substituting the above formula into the expression P(Y|u,i,e) yields:

[0033]

[0034] in: Indicates to The expected value is a constant when the recommendation system ranks the prediction results. It will not change the sorting result; the comparison mainly involves ELU'(f) θ The (u,i) part is used, so ELU'(f) is approximately sorted. θ (u,i)) is regarded as P(Ydo(u,i)), which eliminates the project exposure bias in the project representation.

[0035] The positive effects that this invention can produce are:

[0036] 1. This invention uses a causal inference method to explore the potential mechanism leading to the echo chamber phenomenon, clarifies how variables in the recommendation system influence each other and the mechanism that causes changes in recommendation results, and proposes a dual-representation debiasing recommendation algorithm to effectively alleviate the echo chamber problem in the recommendation system.

[0037] 2. Existing methods use association modeling to build recommendation systems, which cannot explain the effectiveness of existing methods in alleviating the echo chamber problem. At the same time, they are affected by spurious correlations. This invention starts from the perspective of causality, clarifies the mechanism of mutual influence between variables and the change of results, discovers the real cause of the echo chamber problem, removes the confounding factors between user representation, item representation and prediction results, and further reduces the impact of the echo chamber problem on the recommendation system. Attached Figure Description

[0038] Figure 1 This is the cause-effect graph of the recommendation system of the present invention;

[0039] Figure 2 This is a diagram of the recommendation algorithm framework of the present invention. Detailed Implementation

[0040] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0041] This invention proposes a recommendation algorithm based on causal inference to alleviate the echo chamber problem. The main idea is to eliminate the spurious correlation between user representation, item representation and prediction results, that is, to eliminate the erroneous causal relationship caused by bias, so that the prediction results are only affected by the direct causal influence of user representation and item representation.

[0042] Specifically, in response to the spurious correlation between project representation and model prediction results, this invention first analyzes the causal relationship using a causal graph, finding that the spurious correlation is caused by project exposure bias due to historical imbalance distribution. To eliminate the erroneous causal relationship between project representation and prediction results caused by the exposure of the confusion factor, this invention employs a backdoor adjustment method to avoid the impact of exposure bias.

[0043] To address the spurious correlation between user representations and model predictions, this invention first analyzes the causal relationship using a causal graph, revealing that the spurious correlation stems from user selection bias caused by historical imbalances. Considering that user representations are determined by user interactions, which continuously generate new interactions in a high-dimensional dynamic space, and that the impact of interactions on user representations changes with the model and training process, this invention employs counterfactual reasoning. By hypothesizing "If a user dislikes this type of item, what content should the recommendation system recommend?", this approach avoids the bias that leads users to prefer items they've already encountered, thus eliminating the impact of selection bias. This method alleviates the echo chamber problem in recommendation systems.

[0044] First, the underlying mechanisms of the echo chamber problem are analyzed based on the cause-effect graph of the recommender system. (See cause-effect graph for details.) Figure 1 Ideally, the predicted result Y should only be directly influenced by the user representation U and the item representation I, i.e., it should only include the paths U→Y and I→Y. These two paths match user features with item features, resulting in a result based on the user's true preferences. However, in reality, the path from U to Y not only exists as U→Y but also as U←S→Y. This indicates that user choices are biased, leading them to interact more with items they have already expressed preference for. Similarly, the path from I to Y not only exists as I→Y but also as I←E→Y. This indicates that items are biased, with items with higher exposure being further shown to the user. The combination of these two factors results in observed user interactions consisting of more items that have already been interacted with, creating an echo chamber effect. A qualified model should avoid amplifying bias and rely on genuine user preferences for recommendations, rather than being influenced by bias.

[0045] According to causality Figure 1 First, we analyze how to represent the interaction probability given a user u, item i, user selection tendency s, and item exposure e in a traditional recommendation model. We describe this task using a formal language, denoted as P(Y|u,i), and then transform this formula:

[0046]

[0047] The following is a formal description of the biased model P(Y|do(u,i)) after intervention. According to the relevant theory of causal inference, it can be directly derived as follows:

[0048]

[0049] The comparison between the traditional model and the bias-reduced model formula explains why the traditional model contains bias effects and cannot reflect users' true interest in the project, while the bias-reduced model, after intervention, can reflect users' true interest in the project. The fundamental reason lies in the two additional factors.

[0050] Secondly, a method for mitigating the echo chamber problem should be designed based on its underlying mechanism. For example... Figure 2 The overall framework diagram illustrates that for user embeddings in the model, counterfactual reasoning is employed to remove user ID representations and user selection representations that are affected by selection bias, retaining only the remaining user feature representations. This aims to remove selection bias while maximizing the utilization of user features. In the user embedding, the green circle represents the user ID embedding p. id The yellow circle represents the user-selected embedded representation p. c The red circle represents the user's remaining feature embedding p. f First, in the counterfactual world—a world where users dislike items they have expressed preference for—we obtain counterfactual user embeddings that do not include selection embeddings. This reduces the impact of selection bias. Since user ID embeddings have a significant impact on recommendation results and are also affected by selection bias, we need to set the degree to which user ID embeddings are removed, represented by dashed circles in the diagram. For item embeddings in the model, a backdoor adjustment method is used, performing causal inference with the aid of a causal graph. Because traditional recommendation models do not quantify the impact of exposure probability on recommendation results, we calculate the exposure probability of each item and then remove the exposure embedding from the item representation using the quantified exposure probability. White circles represent item ID embeddings q. id The blue circle represents the feature embedding q of the project. f The black circle indicates the project's exposure embedding. This represents the exposure of item i at time t. Then, a neural network is trained, and the result after training is the recommendation result after removing user selection bias and item exposure bias.

[0051] Then, counterfactual reasoning is used to remove spurious correlations between user representations and prediction results. This invention argues that users tend to interact with content they like, therefore, most of their historical interaction records contain content they are interested in or like. Content they dislike or are not interested in is often not recorded, leading the model to overlearn user preferences and inducing users to click on content similar to their own preferences. This results in selection bias, where users are more inclined to interact with items similar to their expressed preferences. To reduce the impact of selection bias, the paper uses counterfactual reasoning: "If a user has not expressed a preference for this type of item, then what items should the recommendation system recommend to the user?" Based on this idea, this invention argues that helping users step outside the interest circles of their similar user groups—i.e., echo chambers—and reducing the recommendation of content that user groups are interested in can prevent users from being influenced by selection bias and encourage them to interact with items based on their true preferences, thus avoiding the impact of selection bias.

[0052] For example, a male user with ID 1 selects basketball-related content. His user characteristics include age 30 and gender (male). He belongs to a sports-loving interest group. To avoid selection bias, an intervention P(Y|do(U), I) should be implemented to reduce the number of items his user group prefers. Here, do(U) is formalized as U = (p... id ,do(p c =-g c ), p f ), where g c The user group's preferred content is represented, thus avoiding selection bias caused by the recommendation system catering to users' favorite sports content without changing the rest of the user representation (which reflects the user's true preferences). Furthermore, user interaction acts as a confounding factor, leading to spurious correlations between user IDs and user features. Although user selections can be directly discarded to avoid selection bias, the user ID representation still encodes some content related to those selections, hindering the complete elimination of selection bias. Therefore, while reducing the recommended content for user groups' interests, it is also necessary to reduce the influence of user IDs on the prediction results. For the same reason, this invention uses counterfactual reasoning to imagine "what would the prediction result be if the user representation did not include a user ID representation," described using formal language. The prediction result obtained in the real world is represented as Y. u,i Then the prediction obtained in the counterfactual world is: in The user representation does not include the user ID representation. To obtain the effect of the user ID representation, simply use... Therefore, a weighting factor α can be used to weigh the prediction results after subtracting the influence of the user ID:

[0053]

[0054] Finally, a backdoor adjustment strategy is employed to eliminate spurious correlations between project representations and prediction results. This invention learns parameters based on user interaction history and the BPR loss function used in implicit feedback.

[0055]

[0056] To express P θ 9y=1|u,i,e) can be obtained by simultaneously inputting the user representation $U$, the item representation $I$, and the item exposure representation into the new model. However, considering that integrating user-item matching and exposure together would require reconsidering the integration scheme once the model is changed, it lacks scalability. Therefore, we consider decoupling user-item matching from item exposure:

[0057]

[0058] Considering P θ (y|u,i,e) is a positive value, therefore the model result f needs to be adjusted. θ (u,i) remains positive, therefore a variant of the ELU activation function, ELU', is used, which is specifically expressed as follows:

[0059]

[0060] Regarding the project's exposure Divide user interaction history with the project into time segments and record... Record user interactions with the project at time t. This represents the record of item i being interacted with at time t. The exposure of item i at time t is thus represented as:

[0061]

[0062] in: This represents the number of times item i is observed at time t. It is obtained by summing all user interactions with item i during time t. This represents the sum of the number of interactions with all items within time t.

[0063] This leads to the final training loss function that removes project exposure bias:

[0064]

[0065] Substituting the above formula into the expression P(Y|u,i,e) yields:

[0066]

[0067] in: Indicates to The expected value is a constant, which is the value that the recommendation system assigns when ranking the prediction results. It will not change the sorting result; the comparison mainly involves ELU'(f) θ The (u,i) part can be approximated during sorting as ELU'(f) θ (u,i)) is regarded as P(Y|do(u,i)), and the above content eliminates the project exposure bias in project representation.

[0068] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A recommendation algorithm for mitigating the echo chamber problem based on causal inference, characterized in that: The algorithm eliminates spurious correlations between user representations, project representations, and model prediction results, and eliminates erroneous causal relationships caused by bias. First, it analyzes the causal graph to find that the spurious correlations are caused by project exposure bias due to historical imbalance distribution. Second, it uses a backdoor adjustment method based on counterfactual reasoning to avoid the impact of exposure bias. This paper analyzes how to represent the interaction probability given a user u, item i, user selection tendency s, and item exposure e in a traditional recommendation model. The task is described using a formal language, denoted as P(Y | u, i), and the formula is transformed as follows: ； ； ； ； A formal description of the biased model P(Y | do(u, i)) after intervention is given, and based on the relevant theory of causal inference, the following can be directly derived: ； ； A backdoor adjustment and bias removal method using counterfactual reasoning is employed. For user embeddings in the model, user ID representations and user selection representations affected by selection bias are removed, retaining the remaining user feature representations. This achieves the goal of removing selection bias while utilizing user features. In the user embeddings, green circles represent user ID embeddings. The yellow circle represents the representation selected by the user for embedding. The red circle represents the user's remaining feature embedding. ; First, in the counterfactual world—a world where users dislike items they have expressed preference for—we obtain counterfactual user embeddings that do not include selection embeddings, reducing the impact of selection bias. Since user ID embeddings have a significant impact on recommendation results and are also affected by selection bias, we need to set the degree to which user ID embeddings are removed, represented by dashed circles. For item embeddings in the model, a backdoor adjustment method is used, performing causal inference with the aid of a causal graph. Because traditional recommendation models do not quantify the impact of exposure probability on recommendation results, we calculate the exposure probability of each item and then remove the exposure embedding from the item representation using the quantified exposure probability. White circles represent item ID embeddings. Blue circles represent embedded project features. The black circle indicates the project's exposure embedding. This represents the exposure of item i at time t. Then, a neural network is trained, and the result after training is the recommendation result after removing user selection bias and item exposure bias. Then, counterfactual reasoning is used to remove the spurious correlation between user representations and prediction results. Users tend to interact with content they like, so most of the user's historical interaction records are content that the user is interested in or likes. Content that they dislike or are not interested in is often not recorded. This leads to the model overlearning user preference items, inducing users to click more on content similar to their own preferences. This leads to selection bias, that is, users are more inclined to interact with items that are similar to their expressed preferences. By using counterfactual thinking, "If the user has not expressed a liking for this type of item, then what items should the recommendation system recommend to the user?" Based on this idea, the system helps users step out of the interest circle of their user group with similar interests, i.e., the echo chamber, reducing the recommendation of content that the user group is interested in, avoiding the user being induced by selection, and making them interact with items based on their own true preferences, thus avoiding the impact of selection bias. We use counterfactual reasoning to hypothesize "what would the prediction result be if the user representation did not contain a user ID?" We describe this using formal language, and the prediction result obtained in the real world is represented as follows: Then the prediction obtained in the counterfactual world is: ,in The user representation does not include the user ID representation. To obtain the effect of the user ID representation, simply use... Therefore, a tradeoff factor is used. To weigh the prediction results by subtracting the user ID: ； ； ； Finally, a backdoor adjustment strategy is employed to eliminate the spurious correlation between project representations and prediction results, and the parameters are learned based on the user's interaction history and the BPR loss function used in implicit feedback. ； Want to express This can be achieved by simultaneously inputting user representation $U$, project representation $I$, and project exposure representation into the new model. However, considering that integrating user-project matching and exposure together would require reconsidering the integration scheme if the model is changed, it lacks scalability. Therefore, we consider decoupling user-project matching from project exposure: ； Considering Since the value is positive, the model results need to be evaluated. To maintain a constant positive value, therefore, the following was adopted. Variants of activation functions To achieve this, its specific expression is as follows: ； Regarding the project's exposure The user's interaction history with the project is divided into time segments and recorded. Record user interactions with the project at time t. Indicates in Moment Project The interactive record, thus recording the project at time t. Exposure is expressed as: ； in: Indicates project In time The number observed is determined by time. All users within the project It is obtained by adding the interactions together, that is ; Indicates time Within, the sum of the number of interactions for all items; This leads to the final training loss function that removes project exposure bias: ； Substituting the above formula into the expression P(Y | u, i, e), we get: ； ； ； ； in: Indicates to The expected value is a constant when the recommendation system ranks the prediction results. It will not change the sorting result; the comparison mainly focuses on... Partial, therefore, during sorting Seen as This eliminates project exposure bias in project representation.

2. The recommendation algorithm for mitigating the echo chamber problem based on causal inference according to claim 1, characterized in that: By analyzing the underlying mechanism of the echo chamber problem using causal graphs, the ideal prediction result Y should only be directly affected by the user representation U and the item representation I, that is, it should only contain the paths U → Y and I → Y. These two paths match user characteristics and item characteristics, and the result is derived from the user's true preferences. However, in reality, there is not only a path U → Y from U to Y, but also a path U ← S → Y. This indicates that user choices are influenced by bias and will tend to interact with items that have already expressed preferences. Similarly, there is not only a path from I to Y, but also a path from I to E to Y. This indicates that the project is affected by the bias. Projects with higher exposure will be further shown to users. The combination of the two makes the observed user interaction consist of more interactive projects, i.e., being trapped in the echo chamber. A good model should avoid amplifying biases and rely on genuine user preferences for recommendations, rather than being influenced by biases.

Images