A personal and group-oriented behavior internet modeling and prediction recommendation method

Abstract

The application discloses a kind of individual and group-oriented behavior internet modeling and prediction recommendation method, the method includes the following steps: step 1: establish the behavior preference model of behavior internet based on hypergraph;Step 2: extract the relationship between behaviors based on the model of step 1;Step 3: discover the behavior mode of user based on the behavior relationship of step 2;Step 4: predict the behavior trend of user based on the behavior relationship of step 2 and the behavior mode of step 3;Step 5: personalized service recommendation is carried out for user based on step 2, step 3, step 4.The model of behavior internet is proposed in the application, and the personalized behavior internet is obtained by cause-effect inference and relationship discovery, and the recommendation is carried out by using the deep learning method based on external knowledge, which improves the recommendation efficiency and effect for individual and group, overcomes the shortcomings that the traditional recommendation method based on deep learning is difficult to fully utilize personalized knowledge, and realizes the personalized and efficient recommendation of service.

Description

Technical Field

[0001] This invention belongs to the field of service recommendation and relates to a behavioral internet modeling and prediction recommendation method for individuals and groups. Background Technology

[0002] With the development of computer technology, network technology, and sensing technology, the Internet of Things (IoT) has been widely applied in many fields, and IoT devices (such as sensors, actuators, cameras, etc.) are ubiquitous. User behavior data collected by IoT devices constructs a behavioral internet for individuals and groups, thereby enabling behavior prediction and service recommendations tailored to individual and group users. However, existing methods struggle to effectively utilize personalized data for behavior prediction and service recommendations for different individuals and groups. Summary of the Invention

[0003] To address the shortcomings of traditional deep learning-based recommendation methods in fully utilizing personalized knowledge, this invention provides a behavioral internet modeling and prediction recommendation method for individuals and groups.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A behavioral internet modeling and prediction recommendation method for individuals and groups includes the following steps:

[0006] Step 1: Establish a hypergraph-based behavioral internet behavior preference model N = (B, E), where B represents a node in IoB and E represents the edge between nodes, i.e. the relationship between behaviors;

[0007] If there is a common relationship between two actions, a triple will be formed.<b1,e,b2> This indicates that a relationship e is formed between behavior 1 and behavior 2, and the two behaviors originate from different people i and j, whose behavior networks are N and N, respectively. i and N j Behavior b i1 ∈N i Behavior b j1 ∈N j At this point, the triplet formed is i1 ,e,b j1 >;

[0008] If multiple behaviors have a super relation, an association matrix will be formed, represented by a hypergraph. Store all behavior nodes associated with super-relationships; the corresponding super-relationship association matrix is ​​expressed as follows:

[0009]

[0010] ​Where u is the number of all behavioral nodes in IoB, n is the number of all superrelations in IoB, and if m ij =1 indicates that the super relation j is associated with behavior i; otherwise, the super relation j is not associated with behavior i.

[0011] Step 2: Extract the relationships between behaviors based on the model from Step 1. Specifically, for generating ordinary edges, this is achieved through relation mining. During the learning process, related behaviors are enhanced, while unrelated behaviors are filtered out, resulting in a relation matrix. The elements can take the value 0 or 1, and the matrix representation is as follows:

[0012]

[0013] Among them, T B The types of behaviors involved in an IoB system;

[0014] The relationship matrix Mr is transformed into a graph, which represents the IoB of the user abstract dimension and the relationships between user behavior types;

[0015] Based on the abstract IoB, interpretable machine learning is used to analyze each behavior instance and connect the behavior instances to form a specific IoB;

[0016] Step 3: Discover user behavior patterns based on the behavioral relationships in Step 2. The specific steps are as follows:

[0017] The relationship matrix is ​​updated based on the user at different times. As input, by analyzing the evolution history of the relationship matrix, user behavior patterns over a period of time are categorized into five types according to the following rules:

[0018] 1. If the user's behavioral relationships do not show significant evolution, then it is a stable pattern;

[0019] 2. If a user's behavioral tendencies continue to increase, it is considered an expansion pattern;

[0020] 3. If the user's behavioral correlation tendency continues to decrease, it is in decay mode;

[0021] 4. If user behavior relationships evolve periodically, it is a periodic pattern;

[0022] 5. If the evolution of user behavior relationships does not show a clear pattern, it is a fluctuating pattern.

[0023] The above process is defined as follows:

[0024] M p =C(M) r )

[0025] Where C is a pattern classification algorithm that infers each behavioral relationship m. r Behavioral pattern m p M p This is a behavioral pattern matrix;

[0026] Step 4: Based on the behavioral relationships in Step 2 and the behavioral patterns in Step 3, predict the user's behavioral trends, that is, predict the user's future personalized behavioral relationships. The specific steps are as follows:

[0027] IoB is in a constantly updating state. Each update generates new behavioral nodes and creates new behavioral relationships without disrupting the existing network structure, thus generating a behavioral internet N that includes the user's latest behavioral preferences. t t is the update time, and the snapshot generated by each IoB update is used as the sequence S. N ={N1, N2, ..., N t}, used to predict the behavior relationship matrix at the next time step.

[0028] Step 5: Based on steps 2, 3, and 4, provide personalized service recommendations to the user. The specific steps are as follows:

[0029] Step 51: Utilize behavioral relationships for service recommendations

[0030] The definition of service recommendation based on ordinary relationships is as follows:

[0031]

[0032] Among them, R n For recommendation algorithms, H B s represents the historical behavior of a group of users, and s represents the recommended service, i.e., the service bound to the predicted individual behavior.

[0033] Step 52: Utilize shared behavioral relationships for service recommendation

[0034] The definition of service recommendations for shared relationships within a hyperrelationship is as follows:

[0035]

[0036] Among them, R j For recommendation algorithms, M i This is the hyperrelationship matrix at the current moment;

[0037] Step 53: Utilize behavioral cooperation relationships for service recommendation

[0038] Service recommendation based on cooperation relationships is defined as follows:

[0039]

[0040] Among them, R c In the recommendation algorithm, O represents the goal pursued by the user group.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] This invention proposes a behavioral internet model, which obtains a personalized behavioral internet through causal inference and relationship discovery, and uses deep learning methods based on external knowledge for recommendation, thereby improving the efficiency and effectiveness of recommendations for individuals and groups. It overcomes the shortcomings of traditional deep learning-based recommendation methods that are difficult to fully utilize personalized knowledge, and realizes personalized and efficient service recommendations. Attached Figure Description

[0043] Figure 1 A flowchart for modeling, predicting, and recommending internet behaviors for individuals and groups;

[0044] Figure 2 A schematic diagram of a behavioral internet model in a personal setting;

[0045] Figure 3 A schematic diagram of a behavioral internet model applicable to individuals and groups;

[0046] Figure 4 A diagram illustrating the construction and updating of the behavioral internet;

[0047] Figure 5 A schematic diagram illustrating the abstract and concrete forms of the behavioral internet;

[0048] Figure 6 This is a schematic diagram illustrating the optimization of the sequence recommendation model. Detailed Implementation

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0050] This invention provides a method for behavioral internet modeling and prediction recommendation for individuals and groups, such as... Figure 1 As shown, the method includes the following steps:

[0051] Step 1: Establish a hypergraph-based behavioral internet behavior preference model.

[0052] Psychological research shows that human behavior always follows a logical progression because it is controlled by a series of underlying behavioral patterns. The logic of human behavior often connects multiple behaviors, and uncovering these relationships can improve the efficiency and accuracy of behavior-oriented service recommendations. Furthermore, human behavior is not only related to one's own actions but also forms group relationships with the behaviors of others. Using hypergraphs to represent the behavioral internet can depict both individual and group relationships, thus clearly characterizing these features.

[0053] In this step, the hypergraph-based behavioral internet behavior preference model is defined as follows: N = (B, E), where:

[0054] B = {b1, b2, ..., b} u} represents a node in IoB, where b = {behaviorType, ctx, ctn, s}; behaviorType represents the behavior type; ctx represents the context in which behavior B occurs; ctn represents the content of behavior B, i.e., the service resources and information read or generated by the user; s∈S represents behavior b. i The services that are bound to it.

[0055] E = {e1, e2, ..., e} v} represents an edge between nodes, i.e., a relationship between behaviors. e =<st,proba,ctx> Where st is the specific type of the relation, and its value range is {sequential, causal, context, joint, collaborative}; proba refers to the transition probability between two actions, that is, the probability of a certain action occurring after the current action occurs; ctx refers to the context attribute to which the action is attached, and this attribute exists when st is context.

[0056] The definitions of the above parameters are shown in Table 1:

[0057] Table 1

[0058]

[0059]

[0060] Within the global behavioral internet, subnets can exist, with the smallest granularity being an individual's behavioral internet, such as... Figure 2 N t As shown, user behavior associations only occur within their own IoB scope; however, individual behaviors can be interconnected across the internet through group relationships, such as... Figure 2 N tAs shown, relationships can arise between the actions of multiple users. Ordinary edges can connect two actions, while hyperedges can connect multiple actions, thus forming a set of relationships for the group's IoB. Relationships can exist between two or more actions. Relationships that occur only between two actions include: temporal relationships, causal relationships, and contextual relationships, collectively referred to as ordinary relationships; relationships that can occur between multiple actions are called common relationships and cooperative relationships, collectively referred to as hyperrelationships.

[0061] If there is a common relationship between two actions, a triple will be formed.<b1,e,b2> This indicates that a relationship e is formed between behavior 1 and behavior 2. The two behaviors can originate from different individuals i and j, and their respective behavior networks are N. i and N j Behavior b i1 ∈N i Behavior b j1 ∈N j At this point, the resulting triplet is: i1 ,e,b j1 >. For example Figure 3 As shown, In the example, e3 connects two actions b4 and b6 of the same user i1, while e4 connects two actions b6 and b7 of different users i1 and i2.

[0062] Two or more behaviors can be connected by a hyperedge to form a hyperrelation, such as... Figure 2 As shown, e1 is associated with three actions b1, b2, and b3 of three users. If there are super-relationships between multiple actions, an association matrix represented by a hypergraph will be formed. It stores all the behavior nodes associated with the super-relationships. The corresponding super-relationship matrix can be expressed as:

[0063]

[0064] Where u is the number of all behavioral nodes in IoB, n is the number of all superrelations in IoB, and if m ij =1 indicates that the super relation j is associated with behavior i; otherwise, the super relation j is not associated with behavior i.

[0065] IoB updates the network as new behavioral data is imported. Each update generates new behavioral relationships without disrupting the existing network structure, thus creating a behavioral internet that includes the user's latest behavioral preferences. For example... Figure 4 As shown, and Given the actions at two time points in the internet, the update algorithm input is... and b8 output as ​ Structurally, only compared to There is an additional node b8 and two edges e5 and e6.

[0066] In this way, behavior is characterized by the nodes of the graph, ordinary behavioral relationships are characterized by ordinary edges, and super-behavioral relationships are characterized by super-edges. Different behavior types and behavioral relationships are represented by different forms of nodes and edges. Simultaneously, temporal information is represented by the timeline below, allowing temporal relationships to exist naturally within the IoB without requiring mining. The IoB has many snapshots; by analyzing these snapshots, user behavioral characteristics can be obtained, thereby supporting downstream tasks.

[0067] Step 2: Relationship between model extraction behaviors based on Step 1.

[0068] The occurrence of behavior is related to human intention. Based on temporal or causal relationship inference algorithms, potential behavioral relationships can be discovered by using users' historical behavior records, thereby constructing a user behavior relationship matrix chain to support the mining of user behavior patterns and trends.

[0069] This step is divided into two parts: node generation and edge generation. Regarding node generation and hyper-relationship generation, researchers need to identify them according to specific rules based on specific scenarios. This identification method can be arbitrary; it can be an existing behavior recognition algorithm, or it can be based on simple rules defined by the developer (for example, considering an individual's presence in the classroom during class time as "attending class"). This is not the core content.

[0070] For generating ordinary edges, relation mining is generally used. This process employs a recurrent neural network (RNN) and the Notears algorithm to mine ordinary relations. Its core principle is reinforcement filtering: during the learning process, related behaviors are reinforced while unrelated behaviors are filtered out. The RNN can use a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU). Regarding the Notears algorithm, this process optimizes the structural equation model (SEM) to the following objectives:

[0071]

[0072] When calculating the loss function, constraints are still considered:

[0073]

[0074] The algorithm's input is the behavioral history H generated by an agent over a period of time. B (H) B Encode the data to include behavior type, context, and content (arranged chronologically), and output a relation matrix. The elements can take the values ​​0 or 1, and the matrix can be represented as:

[0075]

[0076] Where T B The relation matrix M represents the types of behaviors involved in an IoB system. r This indicates the relationship between behavior types, such as m ij This indicates the relationship between behavior type i and behavior type j. If the relationship here is causal, then M... ij This indicates the probability that the occurrence of behavior type i will lead to the occurrence of behavior type j.

[0077] Relationship matrix M r Essentially, it's an adjacency matrix, which can be transformed into a graph representing the IoB (Inversion of Business) of the user's abstract dimension, showing the relationships between user behavior types. For example... Figure 5 As shown, the abstract level of IoB describes the potential relationships between different types of user behaviors, focusing on reasoning, where each node represents a type of behavior. The concrete level of IoB explains the associations between behavior instances, focusing on interpretation, where each node represents a behavior.

[0078] Building upon the abstract IoB, we need to analyze each behavioral instance through interpretable machine learning, connecting these instances to form a concrete IoB. This process utilizes the Shapley tool, which can interpret the model's predictions. This method is based on Shapley values ​​from game theory, calculating the contribution of each feature to the prediction result. We then use the feature contributions to determine whether the sample has a sufficiently positive / negative impact on the target prediction. Specifically, we first need to train a sufficiently good machine learning model (with evaluation metrics close to or exceeding state-of-the-art) to ensure the interpretability foundation is relatively correct. Then, we use interpretable methods to explain each input of the model. If 1) the weight value of the input instance at the model output is greater than a certain threshold, and 2) the correlation between the behavioral types represented by the instance is greater than a certain threshold, then a correlation can be considered to exist between this group of instances. After completing this process, we can obtain a concrete IoB based on the abstract IoB.

[0079] Step 3: Discover user behavior patterns based on the behavioral relationships in Step 2.

[0080] User behavior may vary over time, but patterns often emerge, which we mine and characterize as behavioral patterns. These patterns reveal the shifting characteristics of users when engaging in different behaviors over a period. Users exhibit varying degrees of behavioral stability and predictability under different patterns. Behavior is easier to predict and more reliable under stable, expanding, and declining patterns; conversely, behavior is more difficult to predict and less reliable. Therefore, summarizing and discovering behavioral patterns can improve the reliability and accuracy of trend and behavioral predictions.

[0081] In this step, behavioral relationships within IoB are used to discover user behavior patterns. This is based on the updated relationship matrix obtained at different times for each user. As input, by analyzing the evolution history of the relationship matrix, user behavior patterns over a period of time are categorized into five types according to the following rules:

[0082] 1. If the user's behavioral relationships do not show significant evolution, then it is a stable pattern;

[0083] 2. If a user's behavioral tendencies continue to increase, it is considered an expansion pattern;

[0084] 3. If the user's behavioral correlation tendency continues to decrease, it is in decay mode;

[0085] 4. If user behavior relationships evolve periodically, it is a periodic pattern;

[0086] 5. If the evolution of user behavior relationships does not show a clear pattern, it is a fluctuating pattern.

[0087] The above process can be defined as:

[0088] M p =C(M) r )

[0089] Where C is a pattern classification algorithm, which infers the behavior pattern m for each behavior relationship mr. p ; The behavior pattern matrix can be represented as:

[0090]

[0091] matrix element m p ∈{stable, expanding, decaying, periodic, fluctuating}.

[0092] The pattern classification algorithm C in this process uses the autocorrelation function to determine periodicity and employs a trend detection algorithm based on the autocorrelation function (ACF) and the Mann-Kendall algorithm, respectively, to detect periodicity and trend. The algorithm's input consists of a relation matrix sequence and parameters such as the effective period of the ACF, confidence threshold, and stationarity tolerance. The algorithm outputs the behavioral patterns of the input sequence.

[0093] Step 4: Based on the behavioral relationships in Step 2 and the behavioral patterns in Step 3, predict the user's behavioral trends, that is, predict the user's future personalized behavioral relationships.

[0094] Based on behavioral relationships, a sequence consisting of behavioral relationship matrices at multiple update times is established to obtain a set of user behavioral feature change histories. This allows for the predictive construction of user behavioral features at the next time step, supporting user behavior prediction and service recommendation.

[0095] like Figure 4 As shown, IoB is in a constantly updating state. Each update generates new behavioral nodes and creates new behavioral relationships without disrupting the existing network structure, thus generating a behavioral internet N that contains the user's latest behavioral preferences. t t represents the update time. Each snapshot generated during an IoB update can be used as a sequence S. N ={N1, N2, ..., N t}, used to predict the behavior relationship matrix at the next time step.

[0096] IoB is essentially a heterogeneous graph, therefore this step can be accomplished using graph evolution prediction algorithms. To adaptively capture the spatial relationships between variables from time-series data, this process employs the following steps for learning:

[0097] (1) First, the IoB nodes are processed using HAN (Heterogeneous Graph Attention Network) to obtain their feature representations:

[0098] v i =HAN(N, i)

[0099] Among them, v i Let i be the feature representation of node i, and N be the IoB network.

[0100] (2) Make f(v) i v j ) represents the link score between behavior node i and node j, and f can be expressed by the following formula:

[0101]

[0102] (3) Obtain the connection probability using the Sigmoid function:

[0103] p(i,j)=Sigmoid(f(v i v j ))

[0104] Where p(i,j) represents the probability that there is a relationship between nodes i and j.

[0105] Step 5: Based on steps 2, 3, and 4, provide personalized service recommendations to users.

[0106] User behavior is highly complex and personalized. To make recommendations more accurate while considering personalized recommendations, it is necessary to comprehensively consider behavioral relationships, behavioral patterns, and behavioral trends.

[0107] In this step, service recommendations are made by leveraging user behavior relationships, behavioral trends, and behavioral patterns. The specific steps are as follows:

[0108] Step 51: Utilize behavioral relationships for service recommendations

[0109] For service recommendations based on ordinary relationships, a modification is made to the structure of a sequence recommendation model by adding a relationship utilization component, enabling it to leverage behavioral relationship trends. To improve recommendation accuracy using external knowledge, service recommendations based on ordinary relationships can be defined as follows:

[0110]

[0111] Among them, R n For recommendation algorithms, H B s represents the user's historical behavior, and s represents the recommended service (i.e., the service associated with the predicted individual behavior).

[0112] Taking BERT4Rec as an example, the specific steps for service recommendation using behavioral relationships are as follows:

[0113] Step 511: Apply certain rules to M r and M p Perform Hadamard product to ensure that each behavioral relationship conforms to its corresponding behavioral pattern:

[0114] M r =M r ⊙map(M P )

[0115] Among them, map is a mapping method that can transform different patterns into different weights.

[0116] Step 512: This process is based on the BERT4Rec algorithm and improves its structure. In the original BERT4Rec model, the output of the last Transformer layer at time t is then passed through a mapping layer to obtain the prediction result. However, this does not take into account the influence of the prediction results at previous times on the prediction result at time t, because the two causal actions span an uncertain time. Therefore, this process adds a behavior relationship utilization layer after the last Transformer layer. The output of the Transformer layer from time 1 to time t is used as input. Based on the prediction results of the target item at these t times, the causal value in the causal matrix Mr is selected and multiplied with the original output to enhance the input with behavior relationship and filter out the input without behavior relationship. Finally, the output of the behavior relationship utilization layer is used as the input of the mapping layer. Specific improvements are as follows: Figure 6 As shown, the computational logic in the right-hand box of the model utilizing behavioral relationships is as follows:

[0117]

[0118] That is, the original input plus itself, the product of the time decay coefficient and the relationship index. The time decay coefficient can be arbitrarily specified, as long as it meets the requirements and conforms to the principle that the longer the behavior is, the smaller the impact on the current behavior. This process adopts a fractional decay coefficient, and the degree of influence decreases linearly.

[0119] Step 52: Utilize shared behavioral relationships for service recommendation

[0120] Service recommendations for shared relationships within a hyperrelationship can be defined as follows:

[0121]

[0122] Among them, R j For recommendation algorithms, H B For the historical behavior of a group of users, s represents the recommended service, and M represents the historical behavior of a group of users. i This is the hyperrelationship matrix at the current moment.

[0123] The specific steps for service recommendation based on shared behavioral relationships are as follows:

[0124] Step 521: Via H B and M i Identify all the historical common relationships involved in this group.

[0125] Step 522: Find (sub)groups that frequently appear in the same super-relationship in the historical common relationship set, and consider them to be users with similar tendencies in common behavior.

[0126] Step 523: Recommend users based on similar user behaviors to the target user.

[0127] This process uses collaborative filtering to handle common relationships. By analyzing users' behavior history, behaviors and users are clustered into several categories. Recommendations are made by selecting possible user behaviors from each category, instead of directly filtering from a large number of behaviors, thus narrowing the target scope and simplifying the calculation.

[0128] Step 53: Utilize behavioral cooperation relationships for service recommendation

[0129] Recommendation based on cooperation relationships is an optimization problem. Service recommendation based on behavioral cooperation relationships can be defined as follows:

[0130]

[0131] Among them, R c For recommendation algorithms, H B O represents the historical behavior of a user group, s represents the goals pursued by the user group, and M represents the recommended services. i This is the hyperrelationship matrix at the current moment.

[0132] Based on the target group and target needs, a group recommendation scheme is designed to meet the maximization / minimization objectives, and the group behavior in the scheme is broken down into individual behavior and recommended to the target user group.

[0133] This process uses the GitHub issue / pull request resolution workflow as an example, considering the collaborative problem-solving process among multiple developers as a cooperative action, with different actions reflecting potential responsibilities. Therefore, we set the goal of group behavior as whether the pull request was successfully merged. By analyzing historical data, we can determine:

[0134] 1. By analyzing the different behaviors of different developers in relation to pull requests, we can infer the behaviors / responsibilities that each developer excels at and assign corresponding weights to them.

[0135]

[0136] Among them, w b V represents the developer's proficiency with behavior b. b It represents all pull requests in which the developer participates with behavior b, suc represents the number of pull requests successfully merged, and α is the factor calculation algorithm, which can be based on V. b Adjust the overall weights, for example, V b The fewer the number, the smaller the regulation factor, and vice versa.

[0137] 2. By analyzing whether different developers successfully participate in pull requests with different tags, we can infer the issue types that each developer is good at handling and assign them corresponding weights.

[0138]

[0139] Here, w t V is based on the developer's proficiency with the tag t. t It represents all pull requests with the tag 't' that the developer participated in, and 'suc' represents the number of pull requests that were successfully merged.

[0140] Based on the data above, we set the optimization objective for problem-solving behavior as maximizing the probability of successfully resolving the issue:

[0141]

[0142] Where p is a probability calculation algorithm based on statistics, s is the parameter in the above formula, and g is the user group to be selected.

[0143] Therefore, when an issue is raised again, by analyzing the above data and maximizing the probability of successful resolution, multiple behaviors can be assigned to multiple people, thereby achieving group behavior recommendation.

[0144] This method can be applied to almost any scenario that supports IoT devices and tracks user behavior. Taking the GitHub issue / pull request process as an example, the following is a recommended example of this method:

[0145] Different developers engage in various behaviors on the GitHub platform, including but not limited to: creating discussions, creating pull requests, adding or removing tags, mentioning, assigning tasks, merging branches, reopening or closing discussions, requesting review, reviewing code, locking, and engaging in discussions.

[0146] For individual developers, each developer has their own independent behavioral network, recording all their actions on the platform. Analysis reveals the relationships between these actions. For example, a developer's "discussion" action always follows the "create discussion" action (based on temporal sequence); a developer's "assign task" action tends to lead to the "merge branch" action (based on causality); and a developer's "add tag" action only leads to the "close discussion" action within a "facebook / react" project (based on context). These behavioral relationships may exhibit different patterns. Analysis reveals these patterns; for example, the "discussion" action always following the "create discussion" action represents a stable pattern, while the "assign task" action tends to lead to the "merge branch" action represents an expanding pattern. Then, based on historical behavioral relationships, the trend of future behavioral relationship changes is predicted. For instance, the probability of the relationship where the "add tag" action only leads to the "close discussion" action within a "facebook / react" project decreases from 0.8 to 0.2, while the probability of the "discussion" action always following the "create discussion" action remains unchanged at 0.75. Finally, the predicted behavioral trends are used to make recommendations to users. For example, if a developer's behavioral sequence in the past 12 hours is: ["Create discussion", "Discussion", "Discussion", "Add tag", "Create PR", "Discussion", "Add tag", "Mention", "Discussion", "Assign task", "Request review", "Discussion", "Merge branch", "Discussion"], the behavioral sequence will be input into the improved model (e.g., a model improved based on BERT4Rec) when making recommendations to the user based on the behavioral relationship utilization module. The model predicts that the user may engage in "relationship discussion" behavior next. Therefore, the IoB system recommends the "Close discussion" service associated with the "Close discussion" behavior to the user.

[0147] In group scenarios, developers can work together as a development team on a project, and the behavior of a group of developers constitutes a group IoB. Suppose there is a group of developers g = {1, 2, 3, 4, 5, ..., 20}, with a total of 20 developers. Each developer has their own IoB, and their IoBs together constitute the group IoB.

[0148] For recommendations based on shared behavioral relationships, services are now recommended to Developer 1 based on these shared relationships. By analyzing the historical behavior of this group, it is found that Developers {2,3,5,7,11,19} and Developer 1 frequently engage in the shared behaviors of "discussion," "review," and "assignment." Therefore, it is believed that they have similarities in these three types of behaviors. Furthermore, Developers {2,3,7,11,19} have been assigned question a (assignment behavior) in the past period. Therefore, based on the shared relationship, the "task acceptance" service, which is associated with the "assignment" behavior, can be recommended to Developer 1.

[0149] For recommendations based on collaborative behavior, the percentage of pull requests (PRs) successfully merged by developers is considered an important factor in measuring developers' problem-solving abilities. A new PR is created, tagged "Style Adjustment" and "Urgent Task." This process requires actions such as "Discussion," "Task Assignment," "Code Review," and "Merging" to complete the PR workflow. Therefore, it is necessary to calculate the problem-solving abilities of 20 developers for this task.

[0150] 1) For tasks tagged "style adjustment", calculate the expertise scores of 20 developers;

[0151] 2) For tasks tagged "urgent task", calculate the proficiency scores of 20 developers;

[0152] 3) Calculate the proficiency scores of 20 developers for behaviors such as "discussion", "assignment", "code review" and "merging".

[0153] The expertise values ​​obtained from processes 1 and 2 are multiplied by the corresponding weights for different behaviors and then summed. Taking the "discussion" behavior as an example:

[0154] y 讨论 =v 讨论 +v 样式调整 *w 讨论,样式调整 +v 紧急任务 *w 讨论,紧急任务

[0155] By calculating the proficiency of four types of behaviors separately, and based on the required number of each behavior, candidate recommenders for the four types of tasks were obtained:

[0156] 1) Discussion: [2,8,16];

[0157] 2) Assigning tasks: [3];

[0158] 3) Review the code: [15,16];

[0159] 4) Merge: [1].

[0160] Therefore, the "Comment" service is recommended for developers 2 and 8, the "Comment" and "Review" services are recommended for developer 16, the "Task Assignment" service is recommended for developer 3, the "Code Review" service is recommended for developer 15, and the "Branch Merge" service is recommended for developer 1.

Claims

1. A personal and group oriented behavioral internet modeling and prediction recommendation method, characterized in that The method comprises the following steps: Step 1: Establishing a hypergraph-based behavior internet behavior preference model wherein: represents a node in the IoB, represents an edge between nodes, i.e., a relationship between behaviors; the , ; represents a behavior type; represents a behavior occurs; represents the content of the behavior occurs, i.e., the service resources and information that the user reads or generates; represents the service to which the behavior is bound, represents the set of services to which the behavior is bound, represents the number of behavior nodes in the IoB system; , wherein, is a specific type of relationship, with a value range of , represents a temporal relationship, represents a causal relationship, represents a contextual relationship, represents a common relationship, represents a cooperative relationship; refers to the transition probability between two behaviors, i.e., the probability of a certain behavior occurring after the current behavior occurs; refers to the context attribute to which the behavior is attached; when is context, the attribute exists; represents the number of behavior relationships in the IoB system; the association matrix is expressed as: wherein, if , indicates a hyper relation , indicates a hyper relation , indicates a hyper relation , indicates a hyper relation , indicates a hyper relation If there is a common relationship between two behaviors, a triple is formed , indicating that behavior 1 and behavior 2 form a relationship e, and the two behaviors come from different people and , whose behavior Internet is and , respectively , behavior At this time, the triple formed is ; If there are hyper relations between multiple behaviors, an association matrix represented by the way of hypergraph is formed , stores all behavior nodes associated with hyper relations, is the number of all behavior nodes in IoB, is the number of all hyper relations in IoB; Step 2: based on the model of step 1, the relationship between behaviors is extracted, the general edge is generated, the related behaviors are enhanced in the learning process, the unrelated behaviors are filtered out, and the relationship matrix is obtained , the value of the element is 0 or 1, is the behavior type involved in the IoB system; the relationship matrix is converted into a graph, which represents the IoB of the user abstract dimension and the relationship between the user behavior types; based on the abstract IoB, the behavior instances are analyzed one by one through the interpretable machine learning, and the behavior instances are connected to form a specific IoB. Step 3: Find the behavior pattern of the user based on the behavior relationship in step 2, and the specific steps are as follows: According to the relationship matrix updated by the user at different times As input, by analyzing the evolution history of the relationship matrix, the user's behavior pattern in a period of time is divided into five categories according to the following rules: 1) If the behavior relationship of the user has no obvious evolution, it is a stable pattern; 2) If the user's behavior association tends to continue to grow, it is an inflation pattern; 3) If the user's behavior association tends to continue to decrease, it is a decay pattern; 4) If the user's behavior relationship evolves periodically, it is a periodic pattern; 5) If the user's behavior relationship evolution has no obvious pattern, it is a fluctuation pattern; The above process is defined as: wherein is a pattern classification algorithm that infers each behavioral relationship behavioral pattern ; is a behavioral pattern matrix; Step 4: Predict the behavior trend of the user based on the behavior relationship in step 2 and the behavior pattern in step 3, that is, predict the future personalized behavior relationship of the user, and the specific steps are as follows: The IoB is in a state of continuous updating, and each time the IoB is updated, a new behavior node is generated, and a new behavior relationship is generated without destroying the existing network structure, so as to generate a behavior internet containing the latest behavior preference of the user , For the updating moment, the snapshot generated each time the IoB is updated is used as a sequence , for predicting the behavior relationship matrix of the next moment ; Step 5: Make personalized service recommendations for the user based on steps 2, 3, and 4, and the specific steps are as follows: Step 51: Use the behavior common relationship to make service recommendations The service recommendation based on the common relationship is defined as follows: wherein, is a recommendation algorithm, is a collective user historical behavior, is a recommended service, i.e. a service to which the predicted individual behavior is bound; Step 52: Use the behavior common relationship to make service recommendations The service recommendation for the common relationship in the super relationship is defined as follows: wherein, is a recommendation algorithm, is a hyper-relational association matrix at the current time instant; Step 53: Use the behavior cooperation relationship to make service recommendations The service recommendation based on the cooperation relationship is defined as follows: wherein, is a recommendation algorithm, is a goal pursued by the user group.

2. The personal and group oriented behavioral internet modeling and prediction recommendation method according to claim 1, characterized in that The specific steps of the step 51 are as follows: Step 511: Hadamard product of and each behavioral relation to the corresponding behavioral pattern: wherein, For the mapping method, different modes can be converted into different weights; Step 512: add a behavior relation utilization layer after the last layer of the BERT4Rec model, taking the outputs of the Transformer layers from time point to time point as inputs, multiplying the causal values in the causal matrix with the original outputs according to the prediction results of the target item at the t time points, enhancing the inputs with behavior relations, filtering out the inputs without behavior relations, and finally taking the output of the behavior relation utilization layer as the input of the mapping layer.

3. The personal and group oriented behavioral internet modeling and prediction recommendation method according to claim 1, characterized in that The specific steps of the step 52 are as follows: Step 521 : finding all the historical common relationship sets involving the group by and finding all the historical common relationship sets involving the group Step 522: Find the group that frequently appears in the same super relationship in the historical common relationship set, and consider them as users with similar tendencies in common behavior; Step 523: Recommend to the user according to the similar user behavior of the target user.

4. The personal and group oriented behavioral internet modeling and prediction recommendation method according to claim 1, characterized in that The specific steps of the step 53 are as follows: According to the target group and target demand, design a group recommendation scheme that satisfies the maximization / minimization target, and split the group behavior in the scheme into individual behavior to recommend to the target user group.

Images