Information cocoon house perception graph convolution recommendation cocoon breaking model

Abstract

The invention discloses an information cocoon house perception graph convolution recommendation cocoon breaking model, which comprises a neighbor sampling module fusing information cocoon house measurement, a negative example enhanced sampling module and an information entropy rearrangement module, and is characterized in that the neighbor sampling module fusing the information cocoon house measurement calculates the information entropy of a user and determines a sampling item set of the user; the negative example enhanced sampling module calculates dimensions similar to user features, decouples the similar dimensions to obtain intensity dimensions and event dimensions in vectors and enhances the event dimensions, and the information entropy rearrangement module calculates embedded vectors of users and items in each layer according to neighborhoods of the users and the items obtained by the neighbor sampling module, and rearranges the embedded vectors of the users and the items in each layer according to the neighborhoods of the users and the items. And the final embedding vector of the user and the project is calculated according to the layer attention module. According to the method, recommendation of different diversified degrees can be provided for users in information cocoon rooms of different degrees, so that formation of the information cocoon rooms is effectively avoided, and the users are assisted to break the information cocoon rooms to a certain degree.

Description

Technical Field

[0001] The present invention relates to a recommendation model, in particular to an information cocoon-aware graph convolution recommendation model. Background Art

[0002] Since American scholar Sunstein proposed the concept of information cocoon, this phenomenon has attracted widespread attention from researchers. The information cocoon phenomenon is described as the narrowing of user information and the polarization of opinions. Existing studies have shown that recommendation algorithms widely used in Internet platforms may accelerate the formation of user information cocoons. Nguyen et al. conducted a statistical analysis on the diversity of recommendation lists in collaborative filtering recommendation models and found that the diversity index showed a significant statistical downward trend during the observation period. Kalimeris et al. confirmed through modeling and analysis of matrix decomposition recommendation models that even in simple matrix decomposition models, there is an amplification phenomenon of user preferences, which further promotes the formation of information cocoons. Piao et al. pointed out through simulation analysis that the two major characteristics of recommendation algorithms are the key to promoting the formation of information cocoons. These two major characteristics are the matching mechanism based on similarity and the imbalance of positive and negative feedback utilization.

[0003] Specifically, the recommendation algorithm in practice uses click-through rate as the optimization goal, deeply explores user preferences by utilizing the user's existing interactive information, and recommends content using a similarity-based matching mechanism. This feature amplifies user preferences in the feedback loop of recommendation-click-re-recommendation, prompting users to enter the information cocoon. Not only that, due to practical factors, users' negative feedback is often very scarce. This leads to an imbalance in the use of positive and negative feedback by the recommendation algorithm, passively emphasizing the characteristics of positive examples, and thus reducing the content diversity of the recommendation list.

[0004] Recommendation algorithms, which are widely used on various Internet platforms, have promoted the formation of information cocoons. The information cocoon effect is manifested in that users are gradually isolated from diverse information and eventually trapped in a single topic or viewpoint. Although information cocoons may give people a sense of security in a comfort zone, they also have potential hazards. Therefore, it is of great significance to improve recommendation algorithms to prevent users from experiencing the information cocoon effect.

[0005] In recent years, recommendation methods based on graph convolution networks (GCNs) have received widespread attention. Graph convolution networks aggregate neighborhood information to obtain node embeddings. Compared with traditional methods, the multi-layer convolution operation of graph convolution methods makes it possible for the model to learn the diverse information of high-order neighbors using the high-order connectivity of the graph. However, recommendation methods based on graph convolution networks have two major characteristics: similarity matching and imbalanced utilization of positive and negative feedback. This is considered to be an important cause of information cocoons. The similarity-based matching mechanism amplifies user preferences in the feedback loop of recommendation-click-re-recommendation, which directly leads to a narrowing of the user's exposure to information. The imbalance in the utilization of positive and negative feedback is subject to objective factors. Users' positive feedback can often be recorded as positive examples as interactive behaviors to participate in user preference modeling, while negative feedback is difficult to capture. In most existing graph convolution recommendation models, the negative feedback information participating in model training is a negative example randomly sampled from a set of items that have never been observed to interact. However, most of the negative examples sampled by this method are easy negative examples that are not similar to positive examples. These negative examples contain little information, making it difficult to guide model training to form a better decision boundary. At the same time, they also make the model pay more attention to the characteristics of positive examples, promoting the formation of information cocoons.

[0006] As for users who are already in information cocoons, previous researchers have taken the pursuit of the highest possible diversity as the optimization goal under the condition of minimizing the loss of recommendation accuracy. This brings about the dilemma of diversity and accuracy. At the same time, according to the theory of cognitive dissonance, when people are impacted by new ideas, they tend to maintain their existing ideas. Therefore, for users with a deeper degree of information cocoon, if the diversity of the recommendation list is too high, it may cause the user's defensive reaction, making them more determined to stick to their original views. Not only that, users in a mild cocoon are relatively less affected by the cocoon. Over-emphasizing the diversity of recommendations for these users will make it impossible for the model to accurately model the user's preferences. At present, recommendation algorithms play an important role in alleviating the problem of information overload. However, the existing recommendation methods promote the emergence of user information cocoon phenomenon when pursuing recommendation accuracy. Summary of the invention

[0007] The purpose of the present invention is to provide an information cocoon-aware graph convolutional recommendation cocoon-breaking model to reduce the situation where users produce the information cocoon effect.

[0008] The object of the invention is achieved in this way:

[0009] An information cocoon-aware graph convolution recommendation model, characterized by comprising:

[0010] The neighbor sampling module that integrates the information cocoon metric is connected to the negative example enhancement sampling module and the information entropy rearrangement module to construct a subgraph with diversity that is suitable for the degree of user cocooning, thereby improving the model's amplification of user preferences from the data end;

[0011] The negative example enhancement sampling module is connected to the neighbor sampling module and the information entropy rearrangement module that integrate the information cocoon metric to construct excellent negative examples and prevent the model from over-learning the characteristics of positive examples and causing the recommendation list to become homogeneous.

[0012] The information entropy rearrangement module is connected to the neighbor sampling module and the negative example enhancement sampling module that integrates the information cocoon metric, and is used to ensure that the recommendation list is suitable for the user's cocoon level at the recommendation end.

[0013] Furthermore, the neighbor sampling module of the fusion information cocoon metric calculates the user's information entropy, and the calculation formula is:

[0014]

[0015] Among them, i is the neighbor project node of user u, p(i) is the probability of the category to which project i belongs appearing in the neighbors of user u, and H(u) has an upper bound of logn, where n is the total number of categories of the neighbor project nodes of user u.

[0016] Furthermore, the neighbor sampling module of the fusion information cocoon metric maximizes the selected item set S u The total similarity with the unselected item i is calculated as:

[0017]

[0018] Among them, S u is the set of selected items, N u is the set of items to be selected, sim(i,i′) is the similarity between item i and item i′, σ 2 is the Gaussian kernel width.

[0019] Furthermore, the neighbor sampling module of the fusion information cocoon metric adopts a greedy idea to select items from the to-be-sampled item set and add them to the sampling item set.

[0020] Furthermore, the negative example enhancement sampling module decouples the dimensions similar to the user features in the negative example features through the gating module, and the calculation formula is:

[0021] gate hard =σ(W item e n ⊙W user e u )

[0022] Among them, the Sigmoid function σ(·) is used to map the value range to (0,1), W item With W user is the linear transformation of the project and user vectors in the common feature space, ⊙ is the corresponding element-wise product operation, representing e i and e u The degree of similarity in the corresponding dimension;

[0023] The decoupled features are and in,

[0024] Furthermore, the negative example enhancement sampling module enhances the decoupled features, and the calculation formula is:

[0025]

[0026] Among them, the sgn() function is used to smooth the positive example e p With negative example e n The difference in the easy dimension is to avoid introducing too much positive information, which may cause errors in the model; the generated embedding to be sampled From the easy dimension to e dir After the direction enhancement Δ, it is merged with the hard dimension; Δ is set to introduce random noise in the interval [0,0.1].

[0027] Furthermore, the information entropy rearrangement module adjusts the recommendation list using the optimization index MMR, and the calculation formula of the optimization index MMR is:

[0028]

[0029] Among them, U is the user embedding vector, D is the recommendation list, S is the set of selected items in R, and λ is the balance parameter between relevance and accuracy.

[0030] Furthermore, the information entropy rearrangement module includes LightGCN and a layer attention module. The LightGCN captures the feature representation of users and items, and the calculation formula is:

[0031]

[0032] in, is the embedding vector of the k-th layer user, is the embedding vector of the k-th layer item, N u is the user obtained by the neighbor sampling module, N i is the neighborhood of the item obtained by the neighbor sampling module.

[0033] The layer attention module calculates the user embedding vector obtained by LightGCN to obtain the final user embedding vector. The calculation formula is:

[0034] The layer attention module calculates the embedding vector of the item obtained by LightGCN to obtain the final item embedding vector. The calculation formula is:

[0035] The present invention firstly uses a neighbor sampling module based on information entropy to construct a diversified subgraph, thereby weakening preference amplification at the data end; then uses an enhanced negative example sampling module to construct high-quality negative examples to participate in training, thereby alleviating the imbalance problem of positive and negative feedback utilization; finally, the information entropy rearrangement module is combined to block preference amplification in the rearrangement stage.

[0036] The present invention uses a neighbor sampling module based on information entropy to build a subgraph with appropriate diversity for the user and applies it to graph convolution aggregation. The enhanced negative sampling module can provide similar but negative hard negative examples for model training, so that the model pays more attention to a small number of items. The information entropy rearrangement module adjusts the recommendation list to adapt to the user's current cocoon level during the rearrangement stage. Balance diversity and accuracy for different user states.

[0037] The present invention proposes three modules, so that the recommendation model can provide recommendations of different degrees of diversity for users in different degrees of information cocoons, effectively avoiding the formation of information cocoons and helping users to break out of information cocoons to a certain extent. The enhanced negative sampling module constructs difficult negative examples that do not exist in the real data set, effectively alleviating the problem of missing negative feedback and weakening the user preference amplification enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart of the recommended example.

[0039] Figure 2 It is a model framework diagram of the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described below.

[0041] At present, recommendation algorithms play an important role in alleviating the problem of information overload. However, the existing recommendation methods promote the emergence of user information cocoon phenomenon when pursuing recommendation accuracy. To explain the research motivation of this invention, Figure 1 A recommendation example is given. The black icons in the dotted box represent users and their interactive items, the red icons are negative examples obtained by sampling, the icons in the solid box represent the list of recommended items given by the recommendation system after modeling user preferences, and the robot represents the recommendation system.

[0042] There are two main reasons why recommendation systems promote the formation of information cocoons: First, the recommendation method's similarity-based matching mechanism continuously amplifies user preferences in the iteration of recommendation-click-recommendation, making the recommendation list gradually tend to be homogenized. Users passively access homogenized content, leading to information cocoons. Specifically, the recommendation system uses observed user-item interactions to model user interests, and calculates similarity scores based on the characteristics of the recommended items and users to give a recommendation list.

[0043] like Figure 1 As shown in the figure, the ball category accounts for a relatively large proportion of the user's existing interaction records. Compared with a small number of items, the model is more inclined to learn the characteristics of the majority of items. Therefore, the majority of items in the recommended items are selected with a greater probability, while the minority of items are overwhelmed. As a result, the proportion of ball categories in the recommended list has increased compared with the interaction records, and the content diversity has decreased. After the list is pushed to the user, the user's interaction is returned to the recommendation system for learning. It is worth noting that the recommendation system puts items with higher similarity scores at the front of the recommendation list, and users will interact with these front-end items rather than back-end items with a greater probability. Therefore, if there is no intervention, the items that the user interacts with newly will show a higher tendency to homogenize. Since new interaction items usually have a higher weight when calculating user preferences, the recommendation system will further increase the proportion of homogenized content when updating user interests based on this. In the long run, users will passively fall into information cocoons. It can be seen that blocking the recommendation algorithm from amplifying user preferences based on the similarity matching mechanism is of great significance in preventing users from forming information cocoons.

[0044] Second, the recommendation system makes unbalanced use of positive and negative user feedback. In reality, positive feedback is often observed in non-rating forms (such as clicks, viewing time, etc.), while negative feedback behaviors (such as not clicking, stopping viewing early) do not necessarily directly indicate that the user does not like it. This implicit feedback situation makes negative feedback even more scarce. Therefore, current research usually uses random sampling methods to construct negative examples to simulate negative feedback. Cai et al.'s research shows that most of the negative examples obtained by random sampling are easy negative examples, which are not similar to positive examples. Using only these negative examples for training is not only difficult to form a better decision boundary, but also causes the model to emphasize the characteristics of the category where the positive examples are located.

[0045] like Figure 1 As shown in the figure, when basketball and cars are used as positive and negative pairs, the model will tend to model the user's interest as a preference for ball games. When basketball and football are used as positive and negative pairs, football, as a negative example similar to basketball, prompts the model to reduce the probability of recommending ball games, thereby weakening the amplification of user preferences. It can be seen that using a suitable negative example sampling method can effectively alleviate the problem of missing negative feedback and provide a feasible path to prevent the formation of information cocoons.

[0046] In a recommendation task, a set of users is represented as U = {u 1 , u 2 …, u m}, and a set of items is represented as I = {i 1 , i 2 …, i n}. Based on this, a user-item bipartite graph G = (N, E) is constructed, where the node set N is the union of users and items, and the edge set E is used to describe the interaction behavior between users and items, that is, when the edge e u,i exists, there is an interaction between node u and node i.

[0047] The present invention aims to recommend a top-k recommendation list with k items to users by learning the user-item bipartite graph. This list needs to meet two major goals as much as possible: being interesting to users and having as diverse recommendation results as possible. Obviously, there is a dilemma between accuracy and diversity. The present invention attempts to customize and balance this dilemma for users in different degrees of information cocoons to help break the cocoons.

[0048] As Figure 2 shown, the information-cocoon-aware graph convolutional recommendation cocoon-breaking model provided by the present invention specifically includes: a neighbor sampling module that fuses information-cocoon metrics, a negative example enhancement sampling module, and an information entropy rearrangement module that are connected in sequence. The neighbor sampling module that fuses information-cocoon metrics is also connected to the information entropy rearrangement module.

[0049] The graph convolutional method can obtain high-order neighbor information through the information aggregation mechanism and has unique advantages in cocoon-breaking. However, all along, restricted by the objective phenomenon that users tend to interact with like-minded people or items of interest, there is a large amount of homogeneous content in the neighborhood. This makes the graph convolutional method amplify user preferences during the process of aggregating neighborhood information, resulting in homogeneous recommendation lists. At the same time, considering that users in different degrees of cocoons have different acceptance levels for the diversity of the recommendation list. The present invention weakens the amplification of user preferences by the model from the data end by performing neighbor sampling that fuses information-cocoon metrics on the original graph data to provide subgraphs with different diversity contents for training and learning.

[0050] The neighbor sampling module that fuses information-cocoon metrics uses information entropy as an index to measure the degree of the information cocoon where the user is located, calculates the information entropy of a specific user u, and the calculation formula is:

[0051]

[0052] where i is the neighbor item node of user u, p(i) is the probability that the category to which item i belongs appears among the neighbors of user u, and H(u) has an upper bound of logn, where n is the total number of categories of the neighbor item nodes of user u.

[0053] The neighbor sampling module integrating information cocoon metric maximizes the selected itemset S u The total similarity with the unselected item i is calculated as:

[0054]

[0055] Among them, S u is the set of selected items, N u is the set of items to be selected, sim(i,i′) is the similarity between item i and item i′, σ 2 is the Gaussian kernel width.

[0056] Through formula (2) and formula (3), it is expected that for each project in the initial project set, there is a similar project in the selected project set.

[0057] Obviously, when there is no limit on the number of sampling times, the optimal solution of the function is to add all the initial item sets to the sampled item sets. The present invention limits the number of sampling times k to n+H(u) and the number of neighboring item nodes |N u | takes a smaller value, and introduces the information cocoon degree parameter to constrain it.

[0058] The neighbor sampling module of the fusion information cocoon metric adopts the greedy idea to select items from the sampled item set to add to the sampled item set. The greedy idea is used to optimize the function value, starting from the empty node flow, and each time the item that can make the function value grow the most is selected from the sampled item set. Add the sampling item set. Its function expression is:

[0059]

[0060] After k calculations, the sampled item set of the user node is obtained and applied to subsequent model calculations.

[0061] For graph convolution methods based on implicit feedback, negative examples are crucial. In most studies, negative examples with no observed interactions are randomly sampled as negative examples to guide model learning. The most advanced idea at present is to use difficult negative examples that carry more effective information to assist the model in forming a better decision boundary. Not only that, compared with easy negative examples, negative pairs consisting of difficult negative examples and positive examples can promote the model to learn the features of positive examples more accurately during model training, effectively avoiding the decrease in diversity caused by learning irrelevant features. However, the sampling method currently commonly used is subject to the objective limitations of sampling in real data sets, and there may be problems such as failure to obtain high-quality difficult negative examples and possible sampling of future positive examples. The present invention constructs difficult negative examples that do not exist in the data set by using a three-step strategy of decoupling-enhancement-sampling, thereby promoting the recommendation diversity of the model.

[0062] First, we randomly sample the unobserved items to obtain the initial negative example set ε, and then pass it through the gating module gate hard Calculate which feature dimensions in the negative examples in ε affect the strength of the negative examples. Here, the strength of the negative examples is defined as the degree of similarity to the user features, that is, the gating module decouples which dimensions in the negative example features are similar to the user features.

[0063] The negative example enhancement sampling module decouples the dimensions of the negative example features that are similar to the user features through the gating module. The calculation formula is:

[0064] gate hard =σ(W item e n ⊙W user e u ) (5)

[0065] Among them, the Sigmoid function σ(·) is used to map the value range to (0,1), W item With W user is the linear transformation of the item and user vector in the common feature space, and ⊙ is the corresponding element-wise product operation, which means e i and e u The degree of similarity in the corresponding dimension.

[0066] The decoupled features are and in,

[0067] After decoupling, we get the intensity dimension and easy dimension in the vector. Since the intensity dimension contains a lot of key information for model training, we will enhance the easy dimension in the enhancement phase. The enhancement step includes two aspects: enhancement in the vector direction and enhancement in size. The calculation formula is:

[0068]

[0069] Among them, the sgn(·) function is used to smooth the positive example e p With negative example e n The difference in the easy dimension is to avoid introducing too much positive information, which may cause errors in the model; the generated embedding to be sampled From the easy dimension to e dir After the direction enhancement Δ, it is merged with the hard dimension; Δ is set to introduce random noise in the interval [0,0.1].

[0070] In the sampling stage, in order to select appropriate hard negative examples from the enhanced candidate negative example set, the score of the candidate negative example is obtained by considering the score and the score after amplification at the same time and balancing it with the hyperparameter ε. The calculation formula is:

[0071]

[0072] Finally, when using the enhanced negative sampling method, the model parameters are optimized by the loss function L. Its calculation method is shown in formula (9). 2 is a regular hyperparameter, and γ is a hyperparameter for adjusting contrast loss and decoupling loss.

[0073]

[0074] The recommendation method uses a similarity matching mechanism to continuously amplify user preferences in the iteration of recommendation-click-re-recommendation. In order to effectively block the preference amplification effect of this mechanism, the accuracy and diversity of recommendations are dynamically balanced for users in different levels of cocoons. The present invention uses an information entropy rearrangement module to adjust the recommendation list using the optimization index MMR. The calculation formula of the optimization index MMR is:

[0075]

[0076] Among them, U is the user embedding vector, D is the recommendation list, S is the set of selected items in R, and λ is the balance parameter between relevance and accuracy.

[0077] As a deep learning method for non-Euclidean data, graph convolutional networks have been applied to many fields including recommendation systems, social networks, etc. Representative works such as GCN, GraphSAG, and GAT follow the message passing paradigm, update node embeddings based on neighborhood information, and show excellent performance in tasks such as node classification and link prediction.

[0078] The present invention focuses on the GCN model in the recommendation scenario. In this scenario, the model learns user and item embeddings based on a bipartite graph consisting of users and their historical interaction items, and generates a recommendation list based on this. NGCF propagates the embeddings of users and items on the graph according to the message passing paradigm, effectively introducing collaborative information explicitly into the embedding generation process. He et al. proposed the LightGCN model by removing feature transformation and linear activation, achieving a dual improvement in training speed and recommendation accuracy. Researchers have proposed many improvements based on LightGCN. Peng et al. further simplified LightGCN and improved recommendation accuracy by using the truncated singular value decomposition (SVD) method. Huang et al.'s research shows that in item sampling and model training, negative examples can show features similar to positive examples. They proposed the SiGRec model to encode positive and negative feedback information to learn the embeddings of users and items, which effectively improved the recommendation accuracy compared to the LightGCN model. Zheng et al. proposed sampling negative examples from items that are similar to positive examples but negative to participate in training. Experiments show that these negative examples significantly improve the diversity of the recommendation list.

[0079] Although the above studies have shown relatively good performance, they have not proposed a better solution to the problem of amplifying user preferences and the imbalance of positive and negative feedback utilization. In current research, mainstream research based on graph convolutional networks still takes recommendation accuracy as the optimization goal, which inevitably leads to the amplification of user preferences. Although some researchers have tried to use negative feedback information to improve the recommendation diversity of the model, their sampling methods have the defect that future positive examples may be sampled and reduce model performance. The present invention aims to improve the above problems by using an enhanced negative sampling method to generate excellent negative examples that do not exist in the data set, and dynamically balance the dilemma of diversity and accuracy for users in different degrees of cocoons.

[0080] The information entropy rearrangement module is optimized with a greedy mindset. First, the item with the highest similarity to the user embedding is selected to be added to the selected set, and then in each iteration, the item with high similarity to the user but low maximum similarity to the items in the selected set is selected to be added to the selected set.

[0081] Users in a mild cocoon are relatively less affected by the cocoon. Overemphasizing the diversity of recommendations for these users will make it impossible for the model to accurately model user preferences. For users with a deeper degree of information cocooning, if the diversity of the recommendation list is too high, it may cause a defensive reaction in the user, making him more determined to stick to his original point of view. Therefore, the present invention sets the balance parameter λ to 0 for users in a mild cocoon, sets the balance parameter λ to x for users in a deep cocoon, and sets the balance parameter λ to y for the remaining users.

[0082] The information entropy rearrangement module includes LightGCN and layer attention module. Based on the previous graph convolution method, LightGCN achieves faster training speed and improved recommendation accuracy by removing feature conversion and nonlinear activation process. This model can quickly and effectively capture the feature representation of users and items. LightGCN captures the feature representation of users and items, and the calculation formula is:

[0083]

[0084] in, is the embedding vector of the k-th layer user, is the embedding vector of the k-th layer item, N u is the user obtained by the neighbor sampling module, N i is the neighborhood of the item obtained by the neighbor sampling module.

[0085] The layer attention module calculates the user embedding vector obtained by LightGCN to obtain the final user embedding vector. The calculation formula is:

[0086] The layer attention module calculates the embedding vector of the item obtained by LightGCN to obtain the final item embedding vector. The calculation formula is:

Claims

1. A graph convolution recommendation model based on information cocoon perception, characterized in that: include: The neighbor sampling module that integrates the information cocoon metric is connected to the negative example enhancement sampling module and the information entropy rearrangement module to construct a subgraph with diversity that is suitable for the degree of user cocooning, thereby improving the model's amplification of user preferences from the data end; The negative example enhancement sampling module is connected to the neighbor sampling module and the information entropy rearrangement module that integrate the information cocoon metric to construct excellent negative examples and prevent the model from over-learning the characteristics of positive examples and causing the recommendation list to become homogeneous. as well as The information entropy rearrangement module is connected to the neighbor sampling module and the negative example enhancement sampling module that integrates the information cocoon metric, and is used to ensure that the recommendation list is suitable for the user's cocoon level at the recommendation end.

2. The information cocoon-aware graph convolution recommendation model according to claim 1 is characterized in that: The neighbor sampling module that integrates the information cocoon metric calculates the user's information entropy using the following formula: Among them, i is the neighbor project node of user u, p(i) is the probability of the category to which project i belongs appearing in the neighbors of user u, and H(u) has an upper bound of logn, where n is the total number of categories of the neighbor project nodes of user u.

3. The information cocoon-aware graph convolution recommendation cocoon-breaking model according to claim 2 is characterized in that: The neighbor sampling module that integrates the information cocoon metric maximizes the selected item set S u The total similarity with the unselected item i is calculated as: Among them, S u is the set of selected items, N u is the set of items to be selected, sim(i,i′) is the similarity between item i and item i′, σ 2 is the Gaussian kernel width.

4. The information cocoon-aware graph convolution recommendation cocoon-breaking model according to claim 1 is characterized in that: The neighbor sampling module of the fusion information cocoon measurement adopts a greedy idea to select items from the to-be-sampled item set and add them to the sampling item set.

5. The information cocoon-aware graph convolution recommendation model according to claim 1 is characterized in that: The negative example enhancement sampling module decouples the dimensions of the negative example features that are similar to the user features through the gating module. The calculation formula is: gate hard =σ(W item e n ⊙W user e u ) Among them, the Sigmoid function σ(·) is used to map the value range to (0,1), W item With W user is the linear transformation of the project and user vectors in the common feature space, ⊙ is the corresponding element-wise product operation, representing e i and e u The degree of similarity in the corresponding dimension; The decoupled features are and in, 6. The information cocoon-aware graph convolution recommendation cocoon-breaking model according to claim 5 is characterized in that: The negative example enhancement sampling module enhances the decoupled features, and the calculation formula is: Among them, the sgn(·) function is used to smooth the positive example e p With negative example e n The difference in the easy dimension is to avoid introducing too much positive information, which may cause errors in the model; the generated embedding to be sampled From the easy dimension to e dir After the direction enhancement Δ, it is merged with the hard dimension; Δ is set to introduce random noise in the interval [0,0.1].

7. The information cocoon-aware graph convolution recommendation model according to claim 1 is characterized in that: The information entropy rearrangement module adjusts the recommendation list using the optimization index MMR, and the calculation formula of the optimization index MMR is: Among them, U is the user embedding vector, D is the recommendation list, S is the set of selected items in R, and λ is the balance parameter between relevance and accuracy.

8. The information cocoon-aware graph convolution recommendation model according to claim 2 is characterized in that: The information entropy rearrangement module includes LightGCN and layer attention module. The LightGCN captures the feature representation of users and items. The calculation formula is: in, is the embedding vector of the k-th layer user, is the embedding vector of the k-th layer item, N u is the user obtained by the neighbor sampling module, N i is the neighborhood of the item obtained by the neighbor sampling module; The layer attention module calculates the user embedding vector obtained by LightGCN to obtain the final user embedding vector. The calculation formula is: The layer attention module calculates the embedding vector of the item obtained by LightGCN to obtain the final item embedding vector. The calculation formula is:

Images