Information cocoon house monitoring method and system based on large model user simulation

Abstract

The invention relates to an information cocoon house monitoring method and system based on large model user simulation, and the method comprises the steps: setting positive behavior characteristics and negative behavior characteristics for a group of simulation users, and enabling the simulation users to carry out preset times of interaction with a target recommendation system in sequence; respectively collecting a first recommendation list and a second recommendation list provided by the target recommendation system for each simulation user in each interaction; and calculating a first recommendation diversity sum corresponding to the positive behavior characteristics and a second recommendation diversity sum corresponding to the negative behavior characteristics based on the collected first recommendation list and the second recommendation list, and indicating the information cocoon room degree of the target recommendation system by using the ratio of the first recommendation diversity sum and the second recommendation diversity sum. The solution can accurately quantify the degree of the information cocoon house caused by the recommendation model, thereby providing technical support for subsequent optimization of the recommendation model and breaking of the information barrier.

Description

TECHNICAL FIELD

[0001] The present application relates to personalized recommendation technology, and in particular to a method and system for monitoring the echo chamber effect of a recommendation system. BACKGROUND

[0002] At present, recommendation systems have become an indispensable part of online platforms and are widely used in various scenarios such as e-commerce, social media, news information, short video and music platforms. They can filter personalized content from massive amounts of information to help users cope with information overload. A recommendation system is essentially an information filtering system, and is a process of cyclic interaction between user historical data, recommendation models and user interaction behaviors. The recommendation model is the core module, which can learn the user's interests and hobbies based on user attributes and historical behavior records, and predict the user's preference for a given thing, content, item or service, and make relevant recommendations. The recommendation system can also record the user's interaction behavior with the recommendation results, and then use it as feedback to further train or incrementally update the recommendation model, so as to provide more accurate recommendation results for users using the updated recommendation model. This interest-based, intelligent, personalized and accurate push is now a common media dissemination mechanism.

[0003] However, this "interest-based" personalized recommendation mechanism also has potential negative effects. The recommendation model continuously reinforces the user's existing interests and opinions in the long-term interaction process, leading to a gradual narrowing of the types of information and the range of opinions received by the user, forming the so-called "information echo chamber" phenomenon. In the "information echo chamber", users repeatedly come into contact with similar content, which can gradually limit their worldview, amplify biases, and exacerbate social polarization. However, there is currently a lack of effective methods to accurately analyze or assess the severity of the information echo chamber of different recommendation models. SUMMARY

[0004] To solve the above problems, the present application provides an information echo chamber monitoring scheme based on large model user simulation, which can accurately reflect the degree of information echo chamber caused by the recommendation model, thereby providing technical support for subsequent optimization of the recommendation model and breaking down information barriers.

[0005] According to a first aspect of embodiments of the present application, a cocoon monitoring method is provided, which includes setting positive behavior features for a group of simulated users and making them interact with a target recommendation system for a predetermined number of times. Collecting first recommendation lists provided by the target recommendation system for each simulated user in each interaction, calculating a first recommendation diversity sum of all simulated users based on the collected first recommendation lists. Setting negative behavior features for the group of simulated users and making them interact with the target recommendation system for the predetermined number of times. Collecting second recommendation lists provided by the target recommendation system for each simulated user in each interaction, calculating a second recommendation diversity sum of all simulated users based on the collected second recommendation lists. Using the ratio of the first recommendation diversity sum and the second recommendation diversity sum to indicate the cocoon degree of the target recommendation system.

[0006] In some embodiments, the simulated users are generated based on a large language model, and wherein the setting positive behavior features for the simulated users includes providing an indication to make the simulated users make different selections than the last interaction when facing the recommendation list from the recommendation system in each interaction, and the setting negative behavior features for the simulated users includes providing an indication to make the simulated users make similar selections to the last interaction when facing the recommendation list from the recommendation system in each interaction, wherein the indication is provided to the simulated users in the form of a prompt word.

[0007] In some embodiments, calculating the first recommendation diversity sum of all simulated users based on the collected first recommendation lists includes, for each simulated user, merging the first recommendation lists provided for the simulated user in each interaction into a first recommendation history of the simulated user, and counting the number of categories involved in each item in the first recommendation history as a first diversity value of the simulated user. Summing the first diversity values of all simulated users to obtain the first recommendation diversity sum.

[0008] In some embodiments, calculating the second recommendation diversity sum of all simulated users based on the collected second recommendation lists includes, for each simulated user, merging the second recommendation lists provided for the simulated user in each interaction into a second recommendation history of the simulated user, and counting the number of categories involved in each item in the second recommendation history as a second diversity value of the simulated user. Summing the second diversity values of all simulated users to obtain the second recommendation diversity sum.

[0009] In some embodiments, in each of the predetermined number of interactions between the simulated users and the target recommendation system, the process comprises: the target recommendation system first trains or incrementally updates its recommendation model according to the interaction set obtained from the last interaction, and then generates a recommendation list for each simulated user using the updated recommendation model. Each simulated user selects a portion of items from its recommendation list according to the currently set positive behavior feature or negative behavior feature to form the interaction set of the simulated user. After aggregating the interaction sets of all simulated users, the interaction set obtained from the last interaction is merged to obtain the interaction set of the current interaction. The interaction set obtained from the last interaction used by the target recommendation system in the first interaction is a pre-set set.

[0010] According to a second aspect of the embodiments of the present application, an information cocoon monitoring system is provided, comprising a simulated user module and a statistics module, wherein the simulated user module is configured to: set a positive behavior feature for a group of simulated users, and make the group of simulated users interact with a target recommendation system for a predetermined number of times; and set a negative behavior feature for the group of simulated users, and make the group of simulated users interact with the target recommendation system for the predetermined number of times. The statistics module is configured to: collect a first recommendation list provided by the target recommendation system for each simulated user with the positive behavior feature in each interaction, and calculate a first recommendation diversity sum of all simulated users based on the collected first recommendation list. Collect a second recommendation list provided by the target recommendation system for each simulated user with the negative behavior feature in each interaction, and calculate a second recommendation diversity sum of all simulated users based on the collected second recommendation list. Use the ratio of the first recommendation diversity sum to the second recommendation diversity sum to indicate the degree of information cocoon of the target recommendation system.

[0011] In some embodiments, the simulated users are generated based on a large language model, and wherein the setting of the positive behavior feature for the simulated users comprises providing an indication to make the simulated users make different selections from the recommendation list from the recommendation system in each interaction than in the last interaction, and the setting of the negative behavior feature for the simulated users comprises providing an indication to make the simulated users make similar selections from the recommendation list from the recommendation system in each interaction than in the last interaction, wherein the indication is provided to the simulated users in the form of a prompt word.

[0012] In some embodiments, calculating the first recommendation diversity sum of all simulated users based on the collected first recommendation list comprises: for each simulated user, merging the first recommendation lists provided for the simulated user in each interaction into a first recommendation history of the simulated user, and counting the number of categories involved in each item in the first recommendation history as a first diversity value of the simulated user. Summing the first diversity values of all simulated users to obtain the first recommendation diversity sum.

[0013] In some embodiments, calculating the sum of second recommendation diversity for all simulated users based on the collected second recommendation list includes: for each simulated user, merging the second recommendation lists provided to that simulated user in each interaction into that simulated user's second recommendation history, and counting the number of categories involved in each item in the second recommendation history as the simulated user's second diversity value. The second diversity values ​​of all simulated users are then summed to obtain the sum of the second recommendation diversity.

[0014] In some embodiments, during a predetermined number of interactions between a simulated user and a target recommendation system, each interaction process includes: the target recommendation system first trains or incrementally updates its recommendation model based on the interaction set obtained from the previous interaction, and then uses the updated recommendation model to generate a recommendation list for each simulated user. The simulated user selects a subset of items from its recommendation categories based on currently defined positive or negative behavioral characteristics to form the simulated user's interaction set. After aggregating the interaction sets of all simulated users, these sets are merged with the interaction set obtained from the previous interaction to obtain the interaction set for the current interaction. In the first interaction, the interaction set obtained from the previous interaction used by the target recommendation system is a pre-defined set.

[0015] Compared with the prior art, the information cocoon monitoring system and method according to the embodiments of this application can accurately quantify the degree of information cocoon formation without relying on real user data, separate the contribution of user preference modeling from information cocoon, thereby helping to guide the improvement of recommendation models in recommendable systems, break down information barriers, improve information diversity, and reduce social bias and group polarization.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0018] Figure 1 A schematic diagram of an information cocoon monitoring system according to an embodiment of this application is shown.

[0019] Figure 2 A flowchart illustrating an information cocoon monitoring method according to an embodiment of this application is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and specific embodiments.

[0021] Furthermore, unless otherwise specified, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application may be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc. may be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0022] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0023] As mentioned above, the formation mechanism of information cocoons is particularly prominent in recommendation systems. Recommendation systems can learn user preferences through actions such as clicks, browsing, and likes, creating a positive feedback loop. They continuously push content highly similar to users' past interests, but this also leads to increasingly monotonous content being seen by users, limiting their subsequent behavior. The more firmly the information cocoon formed by the recommendation system, the more difficult it is for users to access new things, resulting in a gradual narrowing of the types of information and viewpoints they receive, exacerbating biases and social polarization.

[0024] However, due to factors such as the high cost of acquiring real users and strict privacy protection requirements, there is currently a lack of precise quantitative indicators for the degree of information cocoon formation in recommendation systems. As a result, it is difficult to measure and evaluate the information cocoon effect brought about by different recommendation systems.

[0025] In response, this application provides a solution for monitoring information cocoons based on user simulation using a large language model. This solution can accurately quantify the degree of information cocoon formation without relying on real user data, separate the contribution of user preference modeling from that of information cocoons, objectively evaluate the merits of recommendation models, and provide repeatable and scalable experimental methods while maintaining the privacy of users' personal information. This can provide technical support for subsequent optimization of recommendation models and breaking down information barriers.

[0026] In this application, "simulated users" refers to virtual user roles created, imitated, or simulated using large language models. Large language models (LMMs) are cutting-edge technologies in the field of artificial intelligence. Their core capability is understanding and generating natural language, utilizing deep learning to comprehend and generate human-like language and content. During training, large language models are exposed to massive amounts of text and have internalized the characteristics of various character archetypes. In this application, a simulated user module is implemented based on a large language model to generate a set of simulated users that can efficiently and stably mimic user behavior. When a specific prompt word is received, the simulated user module activates relevant knowledge patterns, enabling it to engage in dialogue with users as a chatbot with different interest distributions and behavioral characteristics. During the interaction, the "user's" behavior, dialogue, needs, and even personality are generated by the simulated user module, not from a real person. Currently, it is widely used as an internal development and testing tool and is mature enough to be used as a test subject.

[0027] Figure 1 An information cocoon monitoring system according to an embodiment of this application is illustrated. The monitoring system includes a simulated user module and a statistics module, both of which can interact with a target recommendation system. Alternatively, the target recommendation system being evaluated can be any existing recommendation system (e.g., the MovieLens recommendation system used for recommending movies).

[0028] The simulated user module can be a chatbot generated by a large language model, such as ChatGPT, trained on massive amounts of text. Through prompting engineering, the simulated user module is inputted with natural text, including but not limited to interest and behavioral characteristics, and behavioral features used to instruct the simulated user to make decisions, so that each simulated user possesses positive and / or negative behavioral characteristics. For example, prompts can be provided to the large language model-based simulated user to make different choices in each interaction when faced with a recommendation list from the recommendation system, thus embodying positive behavioral characteristics. Prompts can also be provided to the large language model-based simulated user to make choices similar to those in the previous interaction when faced with a recommendation list from the recommendation system, thus embodying negative behavioral characteristics. It should be understood that these prompts can be provided to the simulated user in a manner similar to the prompts used in the large language model.

[0029] like Figure 1As shown, a group of simulated users generated by the simulated user module are sequentially configured with positive and negative behavioral characteristics, and then interact with the target recommendation system a predetermined number of times. The target recommendation system, as the object of evaluation in this application, can be any existing recommendation system. In the initial interaction between the simulated user and the recommendation system, the recommendation system can provide the simulated user with a preset or default recommendation list. The simulated user then selects items from the recommendation list based on their currently configured positive or negative behavioral characteristics. Subsequently, the recommendation system can adjust or train its recommendation model based on the interaction set consisting of items selected by all simulated users in the current interaction. Based on this updated recommendation model, the system provides the simulated user with the recommendation list needed for the next interaction.

[0030] In embodiments of this application, simulated users and the target recommendation system can interact a predetermined number of times. Each interaction process includes: the target recommendation system first trains or incrementally updates its recommendation model based on the interaction set obtained from the previous interaction, and then uses the updated recommendation model to generate a recommendation list for each simulated user; the simulated user selects a subset of items from its recommendation categories based on currently set positive or negative behavioral characteristics to form the simulated user's interaction set; after aggregating the interaction sets of all simulated users, it merges them with the interaction set obtained from the previous interaction to obtain the interaction set for the current interaction. In some embodiments, the interaction set obtained from the previous interaction used by the target recommendation system in the first interaction can be a pre-set interaction set. In some embodiments, the interaction set obtained from the previous interaction used by the target recommendation system in the first interaction can be set to empty, and the recommendation system is configured not to update the recommendation model when the interaction set is empty, but instead directly provide the simulated user with a preset default recommendation list.

[0031] During the aforementioned interaction process, the statistics module can collect the first recommendation list provided by the target recommendation system for each simulated user with positive behavioral characteristics and the second recommendation list provided by the target recommendation system for each simulated user with negative behavioral characteristics in each interaction. Then, based on the collected first recommendation lists, the total first recommendation diversity of all simulated users is calculated (which will be combined below). Figure 2 (Detailed introduction); The total diversity of second recommendations for all simulated users is calculated based on the collected second recommendation list; The ratio of the total diversity of first recommendations corresponding to positive behavioral features to the total diversity of second recommendations corresponding to negative behavioral features is used to indicate the information cocoon degree of the target recommendation system, which can also be referred to as escape potential in the following text.

[0032] In some embodiments, calculating the total recommendation diversity of all simulated users based on the collected recommendation list includes: for each simulated user, merging the recommendation list provided to the simulated user in each interaction into the simulated user's recommendation history, and counting the number of categories involved in each item in the recommendation history as the diversity value of the simulated user; summing the diversity values ​​of all simulated users to obtain the total recommendation diversity.

[0033] In some other embodiments, the information cocoon degree of the target recommendation system can also be calculated in the following way: First, the ratio of the sum of the first recommendation diversity corresponding to positive behavior features to the sum of the second recommendation diversity corresponding to negative behavior features in each round of interaction is counted as the escape potential of each round. Then, the escape potential results of all rounds are averaged to obtain the final overall escape potential of the recommendation system to indicate the information cocoon degree of the target recommendation system.

[0034] For example, in the scenario described above, where simulated users are sequentially assigned positive and negative behavioral characteristics and interact with the target recommendation system t times, the recommendation lists provided by the system to the user in each interaction are collected. Based on this, for each round t, the diversity value of the user u with positive behavioral characteristics is calculated. This is reflected in the number of items from different categories in the recommendation list:

[0035]

[0036] in This represents a mapping from an item to its category. For a collection of items, Given the set of all possible categories, we can obtain a sequence of diversity values ​​for users exhibiting positive behavior in each round:

[0037] ;

[0038] Then, using a similar process, the diversity value of user u with negative behavioral characteristics is calculated, and a sequence of the user's diversity values ​​is obtained. .

[0039] Therefore, for the current round t, the estimated escape potential of the recommendation system is the ratio of the sum of the diversity of all users with positive behavioral characteristics to the sum of the diversity of all users with negative behavioral characteristics:

[0040] ;

[0041] Next, the escape potential results from all rounds are averaged to obtain the final overall escape potential of the recommendation system:

[0042] .

[0043] when The larger the value, the stronger the user's ability to escape the information cocoon through active "exploration" behavior in this recommendation system. The smaller the value, the more robust the information cocoon of this recommendation system.

[0044] In the above embodiments, by setting simulated users to positive and negative behavioral characteristics respectively, and simulating different interest expansion tendencies in multiple rounds of iterative interaction, the information contact diversity of users under the same recommendation algorithm under the two different behaviors is obtained, and a quantitative index of the severity of information cocoon is constructed, which can accurately reflect the degree of information cocoon caused by the recommendation algorithm, and provide a basis for evaluating and optimizing the recommendation algorithm.

[0045] Furthermore, this quantitative indicator is obtained by comparing the interaction behavior of simulated users with positive and negative behavioral characteristics with that of the recommender system, thereby distinguishing between the recommender system's modeling of user preferences and the information cocoon effect. By comparing the performance differences of the same recommender system on positive and negative users, the influence of inherent user preferences on information diversity is eliminated, thus accurately analyzing the information cocoon caused by the recommender algorithm. This method can pinpoint the source of the information cocoon problem and avoid mistakenly attributing user preferences to algorithm design.

[0046] Furthermore, in the embodiments of this application, simulated users with differentiated behavioral characteristics are generated through a large language model driven by prompting engineering, overcoming the difficulty of obtaining real data and achieving efficient simulation of multi-turn interactions. Its technical advantage is that it enables information cocoon monitoring in the absence of real user data, providing a repeatable and scalable experimental platform for subsequent research.

[0047] Figure 2 A flowchart illustrating an information cocoon monitoring method according to an embodiment of this application is shown. The method includes: step S1) setting positive behavioral characteristics for a group of simulated users and having them interact with a target recommendation system a predetermined number of times; step S2) collecting a first recommendation list provided by the target recommendation system for each simulated user in each interaction; step S3) calculating a first recommendation diversity sum for all simulated users based on the collected first recommendation list; step S4) setting negative behavioral characteristics for the group of simulated users and having them interact with the target recommendation system the predetermined number of times; step S5) collecting a second recommendation list provided by the target recommendation system for each simulated user in each interaction; step S6) calculating a second recommendation diversity sum for all simulated users based on the collected second recommendation list; and step S7) using the ratio of the first recommendation diversity sum to the second recommendation diversity sum to indicate the degree of information cocooning in the target recommendation system.

[0048] Specifically, in step S1, positive behavioral characteristics are set for a group of simulated users generated by the simulated user module. This simulated user module is generated based on a large language model. By providing prompts such as willingness to try new things or a reluctance to choose the same categories, the simulated users are set to have positive behavioral characteristics, enabling them to make different choices from the previous interaction when faced with a recommendation list from the recommendation system in each interaction. The number of simulated users can be between 100 and 200, or more depending on the computing power of the model platform (e.g., a computer or mobile phone). The more simulated users, the better the simulated interaction effect.

[0049] Then, simulated users interact with the target recommendation system a predetermined number of times (t rounds). The interaction process includes the following steps s11 to s16. First, in step s11, an initial interaction set A0 is constructed using a portion of the interaction data from the publicly available recommendation dataset, serving as the cold start dataset for the recommendation system. For example, in the movie domain, the "MovieLens dataset" can be used as the cold start dataset for a movie recommendation system, and similarly, the "Last.fm dataset" can be used as the cold start dataset for a music recommendation system. These datasets are characterized by being clean, realistic, diverse in scale, and containing complete interaction results. In the initial interaction, the recommendation model in the recommendation system is denoted as R0.

[0050] In step s12, before performing the current t-th round of interaction, the recommendation system considers the interaction set A from the previous round. t-1 Based on this, we learn and update to obtain an updated recommendation model R. t .

[0051] In step s13, the recommender system uses the updated recommender model R t Generate a recommendation list L for each simulated user u∈U (u,t) The list length can be set according to the application scenario. For example, in a movie recommendation system, a recommendation list with a length of 10 and containing several movie genres can be generated.

[0052] In step s14, a simulated user interacts with the recommendation list, selecting a subset of items to form an interaction set S. (u,t) Simulated users exhibiting positive behavioral characteristics tend to choose items in the current interaction that differ from their previous interaction.

[0053] In step s15, the set S of interactions generated by all simulated users in round t is... (u,t) In summary, combined with the previous round of interaction A t-1 Merge to form a new set of interactions A for the next round. t .

[0054] In step s16, s12 to s15 are repeated until the set t-round interaction is completed.

[0055] In step S2, the first recommendation list provided by the target recommendation system for each simulated user in each interaction is collected. In each round of interaction, the recommendation list generated by the target recommendation system for each simulated user with positive behavioral characteristics can be denoted as L. (u,t) Let L represent the recommendation list provided to user u in round t. The complete recommendation history generated by the recommendation system for each simulated user u with positive behavioral characteristics in round t can be denoted as L. u ={L (u,1) ,L (u,2) ,…,L (u,T) These recommendation lists are collected and stored as a first recommendation list set, used for subsequent calculations of the diversity sum and for evaluating the information cocoon degree of the target recommendation system.

[0056] In step S3, the first recommendation diversity sum of all simulated users is calculated based on the collected first recommendation list corresponding to simulated users with positive behavioral characteristics. For example, for each simulated user, the recommendation list provided to the simulated user in each interaction is merged into the simulated user's recommendation history, and the number of categories involved in each item in the recommendation history is counted as the diversity value of the simulated user; the diversity values ​​of all simulated users are summed to obtain the first recommendation diversity sum.

[0057] In step S4, the simulated user is given negative behavioral characteristics and then interacts with the recommendation system t times. This can be achieved by providing prompts such as only selecting items the user is familiar with or tending to choose items from the same category. This sets the simulated user to have negative behavioral characteristics, causing them to make similar choices to the previous interaction when faced with the recommendation list from the recommendation system in each interaction. The interaction process is similar to step S1.

[0058] In step S5, the second recommendation list L' provided by the target recommendation system for each simulated user with negative behavioral characteristics in each interaction is collected. (u,t) These recommendation lists are collected and stored as a second set of recommendation lists for subsequent diversity calculations.

[0059] In step S6, the sum of the second recommendation diversity for all simulated users is calculated based on the collected second recommendation list. The calculation process is similar to that in step S3 and will not be repeated here.

[0060] Finally, in step S7, the ratio of the sum of the first recommendation diversity to the sum of the second recommendation diversity is used to indicate the degree of information cocoon in the target recommendation system.

[0061] In the above embodiments, based on the interaction results of simulated users with different behavioral characteristics with the same recommendation system, the recommendation system can distinguish between user preferences and the information cocoon effect, and eliminate the influence of inherent user preferences on information diversity through comparison.

[0062] The terms “comprising,” “having,” “including,” and “containing” should be interpreted as open-ended terms (i.e., meaning “including but not limited to”) unless otherwise stated. The term “connected” should be interpreted as partially or wholly included, attached to, or joined together, even if other things are present in between. The enumeration of numerical ranges herein is intended only as a shorthand method for individually referring to each individual value falling within a range, unless otherwise stated herein, and each individual value is incorporated into the specification as if it were individually enumerated herein. All methods described herein may be performed in any suitable order unless otherwise stated herein or the context clearly contradicts it. Any and all examples or exemplary wording provided herein (e.g., “for example”) are intended only to better illustrate the implementation of this application and, unless otherwise stated, do not constitute a limitation on the scope of this application. No wording in the specification should be construed as indicating that any unstated element is essential to the practice of this application.

[0063] Although this application has been described through preferred embodiments, it is not limited to the embodiments described herein, and various changes and modifications are included without departing from the scope of this application.

Claims

1. A method for monitoring echo chamber, comprising: setting positive behavior features for a group of simulated users and making them interact with a target recommendation system for a predetermined number of times; collecting first recommendation lists provided by the target recommendation system for each simulated user in each interaction; calculating a first recommendation diversity sum of all simulated users based on the collected first recommendation lists; setting negative behavior features for the group of simulated users and making them interact with the target recommendation system for the predetermined number of times; collecting second recommendation lists provided by the target recommendation system for each simulated user in each interaction; calculating a second recommendation diversity sum of all simulated users based on the collected second recommendation lists; using a ratio of the first recommendation diversity sum and the second recommendation diversity sum to indicate an echo chamber degree of the target recommendation system.

2. The method of claim 1, wherein the simulated users are generated based on a large language model, and wherein, the setting positive behavior features for simulated users comprises providing instructions for simulated users to make different selections from the recommendation lists from the recommendation system in each interaction than in the last interaction, the setting negative behavior features for simulated users comprises providing instructions for simulated users to make similar selections from the recommendation lists from the recommendation system in each interaction than in the last interaction, wherein the instructions are provided to simulated users in the form of prompt words.

3. The method of claim 1, wherein the calculating a first recommendation diversity sum of all simulated users based on the collected first recommendation lists comprises: for each simulated user, merging the first recommendation lists provided for the simulated user in each interaction into a first recommendation history of the simulated user, and counting the number of categories involved in each item in the first recommendation history as a first diversity value of the simulated user; summing up the first diversity values of all simulated users to obtain the first recommendation diversity sum.

4. The method of claim 1, wherein the calculating a second recommendation diversity sum of all simulated users based on the collected second recommendation lists comprises: for each simulated user, merging the second recommendation lists provided for the simulated user in each interaction into a second recommendation history of the simulated user, and counting the number of categories involved in each item in the second recommendation history as a second diversity value of the simulated user; summing up the second diversity values of all simulated users to obtain the second recommendation diversity sum.

5. The method of any one of claims 1-4, wherein the process of each interaction in the predetermined number of interactions of simulated users with the target recommendation system comprises: the target recommendation system first trains or incrementally updates its recommendation model according to the interaction set obtained in the last interaction, and then generates a recommendation list for each simulated user using the updated recommendation model; the simulated user selects a part of items from its recommendation categories to form an interaction set of the simulated user according to the currently set positive behavior features or negative behavior features; after the interaction sets of all simulated users are summarized, they are combined with the interaction set obtained in the last interaction to obtain the interaction set of the current interaction; the target recommendation system first trains or incrementally updates its recommendation model according to the interaction set obtained in the last interaction, and then generates a recommendation list for each simulated user using the updated recommendation model; the simulated user selects a part of items from its recommendation categories to form an interaction set of the simulated user according to the currently set positive behavior features or negative behavior features; after the interaction sets of all simulated users are summarized, they are combined with the interaction set obtained in the last interaction to obtain the interaction set of the current interaction; wherein the interaction set obtained by the target recommendation system in the last interaction is a pre-set interaction set.

6. An echo chamber monitoring system comprising a simulated user module and a statistics module, wherein the simulated user module is configured to: set a group of simulated users with positive behavior characteristics and make them interact with the target recommendation system for a predetermined number of times; and set another group of simulated users with negative behavior characteristics and make them interact with the target recommendation system for the predetermined number of times; the statistics module is configured to: collect the first recommendation list provided by the target recommendation system for each simulated user with positive behavior characteristics in each interaction; calculate the first recommendation diversity sum of all simulated users based on the collected first recommendation list; collect the second recommendation list provided by the target recommendation system for each simulated user with negative behavior characteristics in each interaction; calculate the second recommendation diversity sum of all simulated users based on the collected second recommendation list; use the ratio of the first recommendation diversity sum and the second recommendation diversity sum to indicate the degree of echo chamber of the target recommendation system.

7. The system of claim 6, wherein the simulated users are generated based on a large language model, and wherein, the setting of the simulated users with positive behavior characteristics comprises providing instructions for the simulated users to make different selections from the recommendation list from the recommendation system in each interaction than in the last interaction, the setting of the simulated users with negative behavior characteristics comprises providing instructions for the simulated users to make similar selections from the recommendation list from the recommendation system in each interaction than in the last interaction, wherein the instructions are provided to the simulated users in the form of prompt words.

8. The system of claim 6, wherein the calculating of the first recommendation diversity sum of all simulated users based on the collected first recommendation list comprises: for each simulated user, merging the first recommendation list provided for the simulated user in each interaction into a first recommendation history of the simulated user, and counting the number of categories involved in each item in the first recommendation history as the first diversity value of the simulated user; summing up the first diversity values of all simulated users to obtain the first recommendation diversity sum.

9. The system of claim 6, wherein the calculating of the second recommendation diversity sum of all simulated users based on the collected second recommendation list comprises: for each simulated user, merging the second recommendation list provided for the simulated user in each interaction into a second recommendation history of the simulated user, and counting the number of categories involved in each item in the second recommendation history as the second diversity value of the simulated user; summing up the second diversity values of all simulated users to obtain the second recommendation diversity sum.

10. The system of any one of claims 6-9, wherein each interaction process in the predetermined number of interactions of the simulated users with the target recommendation system comprises: the target recommendation system first trains or incrementally updates its recommendation model according to the interaction set obtained in the last interaction, and then generates a recommendation list for each simulated user using the updated recommendation model; simulate a user to select a part of items from the recommended category according to the positive behavior feature or the negative behavior feature set by the current time to form an interaction set of the simulated user; aggregate the interaction sets of all simulated users, and combine the interaction sets with the interaction set obtained from the last interaction to obtain an interaction set of the current interaction; wherein the interaction set obtained from the last interaction adopted by the target recommendation system in the first interaction is a pre-set set.

Images