Information recommendation method, research report information score prediction model training method and device

Abstract

The application provides an information recommendation method, a training method and device of a research report information scoring prediction model, equipment and a storage medium. It relates to the technical field of machine learning. The method comprises: obtaining a behavior sequence of a target object, and dividing the behavior sequence into multiple session information; performing encoding processing on each session information to obtain a first feature vector of each session information; performing behavior-to-behavior time distance feature extraction on multiple first feature vectors of multiple session information to obtain a second feature vector; performing real intention spatial distance feature extraction on the second feature vector based on a multi-head self-attention mechanism to obtain a preference vector of the target object; and obtaining a research report information recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommending the research report information recommendation list to the target object. The method can improve the accuracy of the recommendation result.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to an information recommendation method, a training method and apparatus for a research report information rating prediction model. Background Technology

[0002] In related technologies, information recommendation for users mainly relies on user browsing or operational information in the corresponding environment to obtain user preferences. This is used to predict the degree of user preference for items that have not been visited and to make recommendations. Common user behavior-based recommendation methods can be divided into two main categories: recommendation based on explicit feedback behavior and recommendation based on implicit feedback behavior.

[0003] Collaborative filtering algorithms, frequently used in explicit feedback-based recommendation systems, primarily rely on mining historical user behavior data to discover user interests and hobbies. Users are then segmented based on these interests, and results with similar or identical interests are recommended. However, these algorithms suffer from issues such as feature vector sparsity and cold start problems. Furthermore, they only model static input item information, ignoring the fact that user interests may change over time, thus failing to capture dynamic shifts in user interests. Consequently, the recommended results are overly simplistic and one-dimensional.

[0004] Implicit feedback-based recommendation algorithms primarily calculate user preferences based on historical behavioral sequences and record metadata about changes in user interests over a period of time, leveraging temporal contextual information. Therefore, compared to collaborative filtering algorithms that only use static features, this approach can improve the accuracy of recommendation results. However, in reality, a user's perceived positive behavior does not necessarily correspond to effective positive outcomes. For example, a user might browse for a long time but not actually find the results interesting. Therefore, the lack of explicit negative examples can lead to excessive noise, negatively impacting the recommendation results. Summary of the Invention

[0005] This application provides an information recommendation method, a training method, apparatus, device, and storage medium for a research report information rating prediction model, which can improve the accuracy of recommendation results and obtain a richer recommendation list.

[0006] In a first aspect, this application provides an information recommendation method, comprising: acquiring a behavioral sequence of a target object and dividing the behavioral sequence into multiple session information; wherein the behavioral sequence includes at least one behavioral information, each behavioral information being behavioral data of the target object in relation to research report information; encoding each session information to obtain a first feature vector for each session information; extracting temporal distance features between behaviors from the multiple first feature vectors of the multiple session information to obtain a second feature vector; extracting spatial distance features of true intent from the second feature vector based on a multi-head self-attention mechanism to obtain a preference vector of the target object; and, based on the preference vector, acquiring a research report information recommendation list for the target object from at least one candidate research report information, and recommending the research report information recommendation list to the target object.

[0007] In one implementation, the step of extracting temporal distance features between behaviors from multiple first feature vectors of the multiple session information to obtain a second feature vector includes: extracting temporal distance features between behaviors from multiple first feature vectors of the multiple session information based on a preset independent time interval gated loop unit to obtain the second feature vector; wherein, the independent time interval gated loop unit is a gated loop unit structure that adds a time gate to store time interval information separately.

[0008] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0009] T n =(e t U t +(ΔtD t )+)

[0010] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0011] w t =(1- t )⊙w t-1 + t ⊙g t

[0012] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r tTo update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n The term "time gate" refers to the effect of a time interval on the independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0013] In one implementation, the step of extracting spatial distance features of the true intent from the second feature vector based on the multi-head self-attention mechanism to obtain the preference vector of the target object includes: obtaining the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector; obtaining the weight ratio of each attention head based on the self-attention value output by each attention head and the second feature vector; and obtaining the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0014] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0015]

[0016] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second feature vector.

[0017] In one optional implementation, obtaining the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head includes: weighted summing of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0018] In one implementation, obtaining the research report recommendation list of the target object from at least one candidate research report based on the preference vector includes: obtaining at least one candidate research report and a vector for each candidate research report; obtaining a predicted score value for each candidate research report based on the preference vector and the vector for each candidate research report; and obtaining the research report recommendation list of the target object from the at least one candidate research report based on the predicted score value for each candidate research report.

[0019] In one alternative implementation, the formula for calculating the predicted score of each candidate research report is as follows:

[0020]

[0021] Among them, y ui f is the predicted score for the i-th candidate research report. u Let e ​​be the preference vector. i Let a be the vector of the i-th candidate research report information. u For the behavior bias term, a i is the bias term for candidate research report information, and W is a pre-trained weight matrix.

[0022] Secondly, this application provides a training method for a research report information rating prediction model, which is applicable to the information recommendation method described in the first aspect. The training method includes: acquiring a behavioral sequence of a sample object and dividing the behavioral sequence into multiple session information; wherein the behavioral sequence includes at least one behavioral information, and each behavioral information is behavioral data of the sample object in relation to research report information; inputting the multiple session information into an initial research report information rating prediction model; wherein the research report information rating prediction model includes an encoding module, a temporal distance feature extraction module, a spatial distance feature extraction module for true intent, and a prediction module; based on the encoding module... Each session information is encoded to obtain a first feature vector for each session information; based on the time distance feature extraction module, the time distance features between behaviors are extracted from the multiple first feature vectors of the multiple session information to obtain a second feature vector; based on the spatial distance feature extraction module of the true intent, a multi-head self-attention mechanism is used to extract the spatial distance features of the true intent from the second feature vector to obtain the preference vector of the sample object; the preference vector is sent to the prediction module to obtain the predicted score value of each candidate research report information, and the research report information score prediction model is trained based on the predicted score value of each candidate research report information.

[0023] In one implementation, the time distance feature extraction module includes an independent time interval gated loop unit; wherein, the calculation process of the independent time interval gated loop unit is as follows:

[0024] T n =(e t U t +(ΔtD t )+)

[0025] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0026] w t =(1- t )⊙w t-1 + t ⊙g t

[0027] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n The term "time gate" refers to the effect of a time interval on the independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0028] In one implementation, the spatial distance feature extraction module based on the true intent employs a multi-head self-attention mechanism to extract the spatial distance features of the true intent from the second feature vector to obtain the preference vector of the sample object. This includes: obtaining the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector through the spatial distance feature extraction module based on the true intent; obtaining the weight ratio of each attention head based on the self-attention value output by each attention head and the second feature vector; and obtaining the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0029] In one implementation, the formula for the prediction module is expressed as follows:

[0030]

[0031] Among them, y ui f is the predicted score value of the i-th candidate research report output by the prediction module. u Let e ​​be the preference vector. i Let a be the vector of the i-th candidate research report information. u For the behavior bias term, a i , where represents the bias term of the candidate research report information, and W is the weight matrix to be trained.

[0032] Thirdly, this application provides an information recommendation device, comprising: a processing module, configured to acquire a behavioral sequence of a target object and divide the behavioral sequence into multiple session information; wherein the behavioral sequence includes at least one behavioral information, and each behavioral information is behavioral data of the target object in response to research report information; an encoding module, configured to encode each session information to obtain a first feature vector of each session information; a first extraction module, configured to extract temporal distance features between behaviors from the multiple first feature vectors of the multiple session information to obtain a second feature vector; a second extraction module, configured to extract spatial distance features of true intent from the second feature vector based on a multi-head self-attention mechanism to obtain a preference vector of the target object; and a recommendation module, configured to acquire a research report information recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommend the research report information recommendation list to the target object.

[0033] In one implementation, the first extraction module is specifically used to: extract the time distance between behaviors from multiple first feature vectors of the multiple session information based on a preset independent time interval gated loop unit, so as to obtain the second feature vector; wherein, the independent time interval gated loop unit is a gated loop unit structure that adds a time gate to store the time interval information separately.

[0034] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0035] T n =(e t U t +(ΔtD t )+)

[0036] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0037] w t =(1- t )⊙w t-1 + t ⊙g t

[0038] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n The term "time gate" refers to the effect of a time interval on the independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0039] In one implementation, the second extraction module is specifically used to: obtain the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector; obtain the weight ratio of each attention head according to the self-attention value output by each attention head and the second feature vector; and obtain the preference vector of the target object according to the self-attention value of each attention head and the weight ratio of each attention head.

[0040] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0041]

[0042] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second feature vector.

[0043] In one optional implementation, the second extraction module is specifically used to: perform a weighted summation of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0044] In one implementation, the recommendation module is specifically used to: obtain at least one candidate research report and a vector for each candidate research report; obtain a predicted score value for each candidate research report based on the preference vector and the vector for each candidate research report; and obtain a recommendation list of research reports for the target object from the at least one candidate research report based on the predicted score value for each candidate research report.

[0045] In one alternative implementation, the formula for calculating the predicted score of each candidate research report is as follows:

[0046]

[0047] Among them, y ui f is the predicted score for the i-th candidate research report. u Let e ​​be the preference vector. i Let a be the vector of the i-th candidate research report information. u For the behavior bias term, a i is the bias term for candidate research report information, and W is a pre-trained weight matrix.

[0048] Fourthly, this application provides a training device for a research report information rating prediction model, which is applicable to the information recommendation method described in the first aspect. The training device includes: a processing module for acquiring a behavioral sequence of a target object and dividing the behavioral sequence into multiple session information; wherein the behavioral sequence includes at least one behavioral information, and each behavioral information is behavioral data of the target object in relation to research report information; an encoding module for encoding each session information to obtain a first feature vector of each session information; a first extraction module for extracting temporal distance features between behaviors from the multiple first feature vectors of the multiple session information to obtain a second feature vector; a second extraction module for extracting spatial distance features of true intent from the second feature vector based on a multi-head self-attention mechanism to obtain a preference vector of the target object; and a recommendation module for acquiring a research report information recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommending the research report information recommendation list to the target object.

[0049] In one implementation, the first extraction module is specifically used to: extract the time distance between behaviors from multiple first feature vectors of the multiple session information based on a preset independent time interval gated loop unit, so as to obtain the second feature vector; wherein, the independent time interval gated loop unit is a gated loop unit structure that adds a time gate to store the time interval information separately.

[0050] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0051] T n =(e t U t +(ΔtD t )+)

[0052] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0053] w t =(1- t )⊙w t-1 + t ⊙g t

[0054] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e tThe hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time intervals of the current action sequence, and ⊙ represents the XOR operator; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0055] In one implementation, the second extraction module is specifically used to: obtain the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector; obtain the weight ratio of each attention head according to the self-attention value output by each attention head and the second feature vector; and obtain the preference vector of the target object according to the self-attention value of each attention head and the weight ratio of each attention head.

[0056] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0057]

[0058] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second feature vector.

[0059] In one optional implementation, the second extraction module is specifically used to: perform a weighted summation of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0060] Fifthly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method as described in the first aspect, or to implement the method as described in the second aspect.

[0061] Sixthly, this application provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in the first aspect, or to implement the method as described in the second aspect.

[0062] In a seventh aspect, this application provides a computer program product, characterized in that it includes a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect, or implements the steps of the method described in the second aspect.

[0063] The information recommendation method, training method, apparatus, device, and storage medium provided in this application can obtain a second feature vector representing temporal distance features based on the behavioral sequence of a target object. Then, based on a multi-head self-attention mechanism, spatial distance features representing the target object's true intent are extracted from the second feature vector to obtain the target object's preference vector. This preference vector is then used to recommend research report information to the target object. This improves the accuracy of the recommendation results and enables the generation of a richer recommendation list. Attached Figure Description

[0064] 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.

[0065] Figure 1 The flow of the information recommendation method provided in the embodiments of this application Figure 1 ;

[0066] Figure 2 The flow of the information recommendation method provided in the embodiments of this application Figure 2 ;

[0067] Figure 3 A schematic diagram of word sequences in a text model provided in an embodiment of this application;

[0068] Figure 4 A schematic diagram of a target object interaction sequence provided in an embodiment of this application;

[0069] Figure 5 A schematic diagram of a gated loop unit network structure provided in an embodiment of this application;

[0070] Figure 6 The flow of the information recommendation method provided in the embodiments of this application Figure 3 ;

[0071] Figure 7 The flow of the information recommendation method provided in the embodiments of this application Figure 4 ;

[0072] Figure 8 A flowchart illustrating a training method for a research report information scoring prediction model provided in this application embodiment;

[0073] Figure 9 A schematic diagram of a target object behavior distance sequence recommendation model provided in an embodiment of this application;

[0074] Figure 10 This is a schematic diagram of the structure of an information recommendation device provided in an embodiment of this application;

[0075] Figure 11 A schematic diagram of the structure of a training device for a research report information scoring prediction model provided in an embodiment of this application;

[0076] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0077] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0078] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0079] In the description of this application, unless otherwise stated, " / " means "or", for example, A / B can mean A or B; "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. The various numerical designations such as "first", "second", etc., involved in this application are only for the convenience of description and are not used to limit the scope of the embodiments of this application, nor do they indicate the order of events.

[0080] Please see Figure 1 , Figure 1 The flow of the information recommendation method provided in the embodiments of this application Figure 1 ,like Figure 1 As shown, the method may include, but is not limited to, the following steps.

[0081] S101: Obtain the behavior sequence of the target object and divide the behavior sequence into multiple session information.

[0082] In the embodiments of this application, the behavior sequence includes at least one piece of behavior information, each piece of behavior information being behavioral data of the target object in relation to the research report information. For example, behavioral data related to interactions with the research report, such as browsing, clicking, viewing, saving, and following.

[0083] For example, the behavior sequence of the target object is obtained, which includes at least one behavior information, and each behavior information is divided into a session information, thereby dividing the behavior sequence into multiple session information.

[0084] S102: Encode each session information to obtain the first feature vector of each session information.

[0085] For example, each session information is encoded using the GloVe (Global Vectors) model to obtain the first feature vector corresponding to each session information.

[0086] S103: Extract the temporal distance feature between behaviors from multiple first feature vectors of multiple session information to obtain a second feature vector.

[0087] For example, based on the time distance between each session information in multiple session information, features are extracted from multiple first feature vectors to obtain second feature vectors.

[0088] S104: Based on the multi-head self-attention mechanism, the spatial distance feature of the true intent is extracted from the second feature vector to obtain the preference vector of the target object.

[0089] For example, based on the multi-head self-attention mechanism, different interest points may be contained behind different operational behaviors of the target object, so as to extract the spatial distance features of the true intention of the second feature vector and obtain the preference vector of the target object.

[0090] S105: Based on the preference vector, obtain a research report recommendation list for the target object from at least one candidate research report information, and recommend the research report recommendation list to the target object.

[0091] For example, based on the preference vector, the target object's preferred research report information is obtained from at least one candidate research report information as a research report information recommendation list, and the research report information recommendation list is recommended to the target object.

[0092] By implementing the embodiments of this application, a second feature vector representing temporal distance features can be obtained based on the behavioral sequence of the target object. Then, based on a multi-head self-attention mechanism, spatial distance features representing the target object's true intentions are extracted from the second feature vector to obtain the target object's preference vector. This preference vector is then used to recommend research information to the target object, thereby improving the accuracy of the recommendation results.

[0093] In one implementation, a second feature vector can be obtained by extracting temporal distance features between multiple first feature vectors using a gated recurrent unit (GRU). See also: Figure 2 , Figure 2 The flow of the information recommendation method provided in the embodiments of this application Figure 2 ,like Figure 2 As shown, the method may include, but is not limited to, the following steps.

[0094] S201: Obtain the behavior sequence of the target object and divide the behavior sequence into multiple session information.

[0095] In the embodiments of this application, step S201 can be implemented in any of the ways described in the various embodiments of this application. The embodiments of this application do not limit this, nor will they be described in detail.

[0096] S202: Encode each session information to obtain the first feature vector of each session information.

[0097] In the embodiments of this application, step S202 can be implemented in any of the ways described in the embodiments of this application. The embodiments of this application do not limit this, nor will they be described in detail.

[0098] S203: Based on a preset independent time interval gated loop unit, extract the temporal distance between behaviors from multiple first feature vectors of multiple session information to obtain a second feature vector.

[0099] In the embodiments of this application, the independent time interval gated loop unit is a gated loop unit structure that incorporates a time gate to separately store time interval information.

[0100] For example, based on a preset independent time interval gated loop unit, the global temporal correlation in the target object behavior sequence is expressed to model the target object behavior, thereby extracting the temporal distance between multiple first feature vectors of multiple session information to obtain a second feature vector.

[0101] Understandably, recommendations based on target object behavior sequences generally employ statistical methods for modeling. However, models obtained through this approach often lack generalization ability and cannot handle complex scenarios. Recurrent neural network (RNN) models and their variants, on the other hand, can preserve prior information and fully retain contextual dependencies during data serialization when processing target object behavior sequences. However, in modeling target object behavior sequences, RNNs and their variants simply treat the information in the behavior sequence as a uniformly distributed, static, sequential sequence, ignoring the potential temporal distances between target object behaviors. These temporal distances reflect the dynamic interests of the target object over a period of time. For an example, please refer to [link to example]. Figure 3 , Figure 3 This is a schematic diagram of word sequences in a text model provided in an embodiment of this application. For example... Figure 3 As shown, D n D represents the vector of the nth word in the text sequence. n-1 and D n These are adjacent words; there is no concept of a time interval between adjacent words in text attributes. See also... Figure 4 , Figure 4 This is a schematic diagram of a target object interaction sequence provided in an embodiment of this application. Figure 4 Chinese K n The report represents the nth interaction with the target object in the sequence of actions, k. n-1 and k n The time interval is the period of interaction between the target and the research report. This time interval is variable. The target's behavior within a shorter time interval may indicate the same interest intention, while the target's interest may have shifted after a longer time interval of target behavior.

[0102] Therefore, this application modifies the traditional gated loop unit by adding a Time-GRU structure that stores time interval information separately through a time gate, in order to express the global temporal correlation in the target object's behavioral sequence, thereby modeling the target object's behavior. For an example, please refer to... Figure 5 , Figure 5 This is a schematic diagram of a gated loop unit network structure provided in an embodiment of this application.

[0103] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0104] T n =(e t U t +(ΔtD t )+)

[0105] g t =anh(ee U g +(w t-1 ⊙z t ⊙T n ) g )+

[0106] w t =(1- t )⊙w t-1 + t ⊙g t

[0107] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n This represents a time gate, which signifies the effect of a time interval on an independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t Furthermore, it assigns greater weight to similar historical behaviors. This differentiates between short-term and long-term interests, enriching user profiles and providing users with more comprehensive recommendation lists.

[0108] It is understandable that the target audience's behavior can be diverse, such as browsing lists, browsing details, browsing duration, and liking actions. Different behaviors also represent the target audience's level of interest to some extent. To improve the accuracy of the results, in some embodiments of this application, greater weight may be assigned to some more decisive behaviors (e.g., liking).

[0109] S204: Based on the multi-head self-attention mechanism, the spatial distance feature of the true intent is extracted from the second feature vector to obtain the preference vector of the target object.

[0110] In the embodiments of this application, step S204 can be implemented in any of the ways described in the various embodiments of this application. The embodiments of this application do not limit this, nor will they be described in detail.

[0111] S205: Based on the preference vector, obtain a research report recommendation list for the target object from at least one candidate research report information, and recommend the research report recommendation list to the target object.

[0112] In the embodiments of this application, step S205 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0113] By implementing the embodiments of this application, a second feature vector representing temporal distance features can be obtained based on the behavioral sequence of the target object and the independent time interval gated recurrent unit. This allows for more accurate acquisition of the user's dynamic interest changes. Furthermore, based on a multi-head self-attention mechanism, spatial distance features representing the target object's true intent are extracted from the second feature vector to obtain the target object's preference vector. This preference vector then enables the recommendation of research information to the target object, thereby improving the accuracy of the recommendation results.

[0114] In one implementation, a multi-head self-attention mechanism and a second feature vector can be used to obtain the weight percentage and output self-attention value of each attention head, thereby obtaining the target object's preference vector based on the self-attention value and weight percentage. For an example, please refer to [link to example]. Figure 6 , Figure 6 The flow of the information recommendation method provided in the embodiments of this application Figure 3 ,like Figure 6 As shown, the method may include, but is not limited to, the following steps.

[0115] S601: Obtain the behavior sequence of the target object and divide the behavior sequence into multiple session information.

[0116] In the embodiments of this application, step S601 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0117] S602: Encode each session information to obtain the first feature vector of each session information.

[0118] In the embodiments of this application, step S602 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0119] S603: Extract the temporal distance feature between behaviors from multiple first feature vectors of multiple session information to obtain a second feature vector.

[0120] In the embodiments of this application, step S603 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0121] S604: Based on the multi-head self-attention mechanism and the second feature vector, obtain the self-attention value output by each attention head.

[0122] Understandably, in a multi-head self-attention mechanism, each attention head needs to learn multiple Q, K, V (query, key, value) matrices independently, which corresponds to learning multiple linearly changing weight matrices of Q, K, and V. The self-attention value is obtained by calculating the dot product of each attention head individually.

[0123] For example, the specific calculation process of the above steps can be represented as follows.

[0124]

[0125] Among them, head i Represents the i-th attention head. R is the set of real numbers. d model d represents the sequence encoding length. k =d v =d model / represents the number of attention heads, 1 ≤ i ≤ 0. As an example, the number of self-attention heads can be set to 8, i.e., = 8. It should be noted that in the embodiments of this application, the second feature vector can be used as a query matrix to more deeply mine the target object's interest tendencies. Then, the output results of the multi-head self-attention are obtained.

[0126] X u =MultiHead(Q,K,V)=Concat(head1,…,head h W o

[0127] Among them, X u For the output result, W o Let be the preset output matrix. The formula for calculating the self-attention value matrix can be expressed as follows.

[0128]

[0129] Therefore, the self-attention value output by each attention head can be obtained. Accordingly, the calculation formula for the i-th attention head can be expressed as follows.

[0130]

[0131] S605: Obtain the weight ratio of each attention head based on the self-attention value and second feature vector output by each attention head.

[0132] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0133]

[0134] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second eigenvector.

[0135] S606: Obtain the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0136] For example, based on the self-attention value and weight ratio of each attention head, the evaluation value corresponding to each attention head is obtained, and the preference vector of the target object is obtained based on the evaluation value corresponding to each attention head.

[0137] In one optional implementation, the preference vector of the target object is obtained based on the self-attention value of each attention head and the weight ratio of each attention head, including: weighted summation of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0138] For example, the self-attention value of each attention head is weighted and summed with its corresponding weight percentage using the following formula.

[0139]

[0140] Among them, f u β represents the preference vector of the target object. k h represents the weight percentage of the k-th attention head. k Let t represent the self-attention value of the k-th attention head, where 1 ≤ k ≤ t, and t is the total number of attention heads.

[0141] S607: Based on the preference vector, obtain a research report recommendation list for the target object from at least one candidate research report information, and recommend the research report recommendation list to the target object.

[0142] In the embodiments of this application, step S607 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0143] By implementing the embodiments of this application, a second feature vector representing temporal distance features can be obtained based on the behavioral sequence of the target object. Based on a multi-head self-attention mechanism and the second feature vector, the weight percentage and output self-attention value of each attention head can be obtained. Then, based on the self-attention value and weight percentage, a preference vector of the target object can be obtained, preserving features relevant to the target object's interests as much as possible. This allows for the recommendation of research information to the target object based on the preference vector, thereby improving the accuracy of the recommendation results.

[0144] In one implementation, a predicted score for each candidate research report can be obtained based on a preference vector, thereby generating a recommendation list of research reports for the target object based on multiple predicted scores. For an example, please refer to [link to example]. Figure 7 , Figure 7 The flow of the information recommendation method provided in the embodiments of this application Figure 4 ,like Figure 7 As shown, the method may include, but is not limited to, the following steps.

[0145] S701: Obtain the behavior sequence of the target object and divide the behavior sequence into multiple session information.

[0146] In the embodiments of this application, step S701 can be implemented in any of the ways described in the embodiments of this application. The embodiments of this application do not limit this, nor will they be described in detail.

[0147] S702: Encode each session information to obtain the first feature vector of each session information.

[0148] In the embodiments of this application, step S702 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0149] S703: Extract the temporal distance feature between behaviors from multiple first feature vectors of multiple session information to obtain a second feature vector.

[0150] In the embodiments of this application, step S703 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0151] S704: Based on a multi-head self-attention mechanism, spatial distance features of the true intent are extracted from the second feature vector to obtain the preference vector of the target object.

[0152] In the embodiments of this application, step S704 can be implemented in any of the embodiments of this application. The embodiments of this application do not limit this and will not be described in detail.

[0153] S705: Obtain at least one candidate research report and a vector for each candidate research report.

[0154] For example, obtain at least one candidate research report information and represent each candidate research report information as a vector to obtain a vector for each candidate research report information.

[0155] S706: Based on the preference vector and the vector of each candidate research report, obtain the predicted score value of each candidate research report.

[0156] For example, the preference vector and the vector of each candidate research report are input into a multilayer perceptron to obtain the predicted score value of each candidate research report through the Sigmoid activation function.

[0157] In one implementation, the formula for calculating the predicted score of each candidate research report can be expressed as follows.

[0158]

[0159] Among them, y ui Let f be the predicted score for the i-th candidate research report. u Let e ​​be the preference vector. i Let a be a vector containing the information of the i-th candidate research report. u For the behavior bias term, a i is the bias term for candidate research report information, and W is a pre-trained weight matrix.

[0160] S707: Based on the predicted score of each candidate research report, obtain a recommended list of research reports for the target object from at least one candidate research report.

[0161] As an example, based on the predicted score of each candidate research report, at least one candidate research report can be sorted from largest to smallest according to the predicted score, and the top few (e.g., the top 3) candidate research reports can be selected as the research report recommendation list for the target object.

[0162] By implementing the embodiments of this application, a second feature vector representing temporal distance features can be obtained based on the behavioral sequence of the target object. Then, based on a multi-head self-attention mechanism, spatial distance features representing the target object's true intentions are extracted from the second feature vector to obtain the target object's preference vector. Finally, a predicted score for each candidate research report is obtained based on the preference vector, thereby generating a recommendation list of research reports for the target object based on multiple predicted scores. This can improve the accuracy of the recommendation results.

[0163] Please see Figure 8 , Figure 8 This is a flowchart illustrating a training method for a research report information scoring prediction model provided in an embodiment of this application. Figure 8 As shown, the method may include, but is not limited to, the following steps.

[0164] S801: Obtain the behavior sequence of the sample object and divide the behavior sequence into multiple session information.

[0165] In the embodiments of this application, the above-mentioned behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the sample object in response to the research report information.

[0166] For example, the behavior sequence of a sample object is obtained, which includes at least one behavior information. Each behavior information is then divided into a session information, thereby dividing the behavior sequence into multiple session information.

[0167] S802: Input multiple session information into the initial research report information scoring prediction model.

[0168] In the embodiments of this application, the research report information scoring prediction model includes an encoding module, a time distance feature extraction module, a spatial distance feature extraction module for true intent, and a prediction module.

[0169] For example, multiple conversational information is input into the encoding module, temporal distance feature extraction module, spatial distance feature extraction module of true intent, and prediction module of the initial research report information rating prediction model.

[0170] In one optional implementation, the time distance feature extraction module includes an independent time interval gated loop unit; wherein the calculation process of the independent time interval gated loop unit is as follows:

[0171] T n =(e t U t +(ΔtD t )+)

[0172] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0173] w t =(1- t )⊙w t-1 + t ⊙g t

[0174] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, a is the preset vector activation parameter, and e tThe hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n This represents a time gate, which signifies the effect of a time interval on an independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0175] S803: The encoding module encodes each session information to obtain the first feature vector of each session information.

[0176] In one embodiment of this application, the encoding module may include a GloVe model, thereby using the GloVe model to encode each session information to obtain a first feature vector corresponding to each session information.

[0177] S804: The temporal distance feature extraction module extracts temporal distance features between behaviors from multiple first feature vectors of multiple session information to obtain a second feature vector.

[0178] For example, based on the time distance feature extraction module, features are extracted from multiple first feature vectors according to the time distance between each session information in multiple session information to obtain second feature vectors.

[0179] S805: Spatial distance feature extraction module based on true intent. It adopts a multi-head self-attention mechanism to extract spatial distance features of true intent from the second feature vector in order to obtain the preference vector of the sample object.

[0180] For example, the spatial distance feature extraction module based on true intent uses a multi-head self-attention mechanism to characterize the different interest points that may be contained behind different operational behaviors of the sample object, so as to extract the spatial distance features of true intent from the second feature vector and obtain the preference vector of the sample object.

[0181] In one optional implementation, the spatial distance feature extraction module based on true intent employs a multi-head self-attention mechanism to extract the spatial distance features of true intent from the second feature vector to obtain the preference vector of the sample object. This includes: obtaining the self-attention value output by each attention head through the spatial distance feature extraction module based on the multi-head self-attention mechanism and the second feature vector; obtaining the weight ratio of each attention head based on the self-attention value output by each attention head and the second feature vector; and obtaining the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0182] S806: Send the preference vector to the prediction module to obtain the predicted score value for each candidate research report, and train the research report score prediction model based on the predicted score value for each candidate research report.

[0183] For example, the preference vector is sent to the prediction module to obtain the predicted score value for each candidate research report. Based on the difference between the predicted score value for each candidate research report and the pre-obtained true score value, the model parameters of the research report information score prediction model are updated to train the research report information score prediction model until the difference between the predicted score value and the true score value is less than a preset difference threshold.

[0184] In one alternative implementation, the prediction module is expressed as follows:

[0185]

[0186] Among them, y ui f is the predicted score of the i-th candidate research report output by the prediction module. u Let e ​​be the preference vector. i Let a be a vector containing the information of the i-th candidate research report. u For the behavior bias term, a i , where represents the bias term of the candidate research report information, and W is the weight matrix to be trained.

[0187] In one optional implementation, the research report information rating prediction model is trained based on the predicted rating value of each candidate research report information, including: using a preset cross-loss function, calculating the cross-loss function of the research report information rating prediction model based on the predicted rating value of each candidate research report information; using the cross-loss function as the objective function, optimizing the objective function using a stochastic gradient descent algorithm, and updating the model parameters of the research report information rating prediction model using an error backpropagation algorithm.

[0188] By implementing the embodiments of this application, a research report information rating prediction model can be trained based on the acquired behavioral sequences of sample objects, enabling the model to more accurately acquire different preferences of target objects' behaviors and generate more accurate recommendation results.

[0189] Please see Figure 9 , Figure 9 This is a schematic diagram of a target object behavior distance sequence recommendation model provided in an embodiment of this application. Figure 9 As shown, this model can segment the behavioral sequences of a target object into different sessions using specific rules, and represent each session as a feature vector using a GloVe model within that session. It also processes different time intervals in the target object's behavioral sequences through a temporal gating unit structure; simultaneously, it uses a multi-head self-attention mechanism to calculate the spatial distance of the target object's intent, thereby obtaining different interest directions of the target object across the same facet. Finally, it obtains the predicted score of the target object for candidate reports through a Sigmoid activation function for recommendation.

[0190] Please see Figure 10 , Figure 10 This is a schematic diagram of the structure of an information recommendation device provided in an embodiment of this application. Figure 10 As shown, the device includes: a processing module 1001, used to acquire a behavior sequence of a target object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the target object in response to the research report information; an encoding module 1002, used to encode each session information to obtain a first feature vector of each session information; a first extraction module 1003, used to extract the temporal distance feature between behaviors from the multiple first feature vectors of the multiple session information to obtain a second feature vector; a second extraction module 1004, used to extract the spatial distance feature of the true intent from the second feature vector based on a multi-head self-attention mechanism to obtain a preference vector of the target object; and a recommendation module 1005, used to obtain a research report information recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommend the research report information recommendation list to the target object.

[0191] In one implementation, the first extraction module 1003 is specifically used to: extract the time distance between behaviors from multiple first feature vectors of multiple session information based on a preset independent time interval gated loop unit to obtain a second feature vector; wherein, the independent time interval gated loop unit is a gated loop unit structure that adds a time gate to store the time interval information separately.

[0192] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0193] T n =(e t U t +(ΔtD t )+)

[0194] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0195] w t =(1- t )⊙w t-1 + t ⊙g t

[0196] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U t U g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n This represents a time gate, which signifies the effect of a time interval on an independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0197] In one implementation, the second extraction module 1004 is specifically used to: obtain the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector; obtain the weight ratio of each attention head based on the self-attention value output by each attention head and the second feature vector; and obtain the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0198] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0199]

[0200] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second eigenvector.

[0201] In one optional implementation, the second extraction module 1004 is specifically used to: perform a weighted summation of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0202] In one implementation, the recommendation module 1005 is specifically used to: obtain at least one candidate research report information and a vector for each candidate research report information; obtain a predicted score value for each candidate research report information based on the preference vector and the vector for each candidate research report information; and obtain a recommendation list of research report information for the target object from the at least one candidate research report information based on the predicted score value for each candidate research report information.

[0203] In one alternative implementation, the formula for calculating the predicted score of each candidate research report is as follows:

[0204]

[0205] Among them, y ui Let f be the predicted score for the i-th candidate research report. u Let e ​​be the preference vector. i Let a be a vector containing the information of the i-th candidate research report. u For the behavior bias term, a i is the bias term for candidate research report information, and W is a pre-trained weight matrix.

[0206] This application provides an information recommendation device that can be used to execute the information recommendation method described in the above embodiments. Its implementation principle and technical effect are similar, and will not be repeated here.

[0207] Please see Figure 11 , Figure 11 This is a schematic diagram of a training device for a research report information scoring prediction model provided in an embodiment of this application. This research report information scoring prediction model is applicable to information recommendation methods as provided in any embodiment of this application. Figure 11As shown, the training device includes: a processing module 1101, used to acquire the behavior sequence of the target object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the target object in response to the research report information; an encoding module 1102, used to encode each session information to obtain a first feature vector of each session information; a first extraction module 1103, used to extract the temporal distance feature between behaviors from the multiple first feature vectors of the multiple session information to obtain a second feature vector; a second extraction module 1104, used to extract the spatial distance feature of the true intent from the second feature vector based on a multi-head self-attention mechanism to obtain the preference vector of the target object; and a recommendation module 1105, used to obtain a research report information recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommend the research report information recommendation list to the target object.

[0208] In one implementation, the first extraction module 1103 is specifically used to: extract the time distance between behaviors from multiple first feature vectors of multiple session information based on a preset independent time interval gated loop unit to obtain a second feature vector; wherein, the independent time interval gated loop unit is a gated loop unit structure that adds a time gate to store the time interval information separately.

[0209] In one alternative implementation, the calculation process of the independent time interval gated loop unit is as follows:

[0210] T n =(e t U t +(ΔtD t )+)

[0211] g t =anh(e t U g +(w t-1 ⊙z t ⊙T n ) g )+

[0212] w t =(1- t )⊙w t-1 + t ⊙g t

[0213] Where σ is the sigmoid function, Δt is the time interval between the action sequence and the current time, and is the preset vector activation parameter, e t The hidden state z is the time interval of the current action sequence. t To reset the door, r t To update the gate, U tU g D t and D g For the preset parameter matrix, g t The candidate hidden states represent the time interval of the current action sequence; ⊙ represents the XOR operator. t D g ∈R, and D t ≤0,T n This represents a time gate, which signifies the effect of a time interval on an independent time interval gated loop unit; where r t and T n Simultaneously controlling the hidden state e of the current action sequence. t And assign a greater weight to historical behaviors that are close to the target.

[0214] In one implementation, the second extraction module 1104 is specifically used to: obtain the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector; obtain the weight ratio of each attention head based on the self-attention value output by each attention head and the second feature vector; and obtain the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head.

[0215] In one alternative implementation, the formula for calculating the weight percentage of each attention head is as follows:

[0216]

[0217] Where, β k Let t represent the weight percentage of the k-th attention head, and t be the total number of attention heads. e is the transpose of the self-attention value output by the k-th attention head. u This is the second eigenvector.

[0218] In one optional implementation, the second extraction module 1104 is specifically used to: perform a weighted summation of the self-attention value of each attention head and its corresponding weight ratio to obtain the preference vector of the target object.

[0219] This application provides a training device for a research report information scoring prediction model, which can be used to execute the technical solution of the training method of the research report information scoring prediction model in the above embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0220] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, processing module 1101 can be a separate processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its functions can be called and executed by a processing element of the above device. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each of the above modules can be completed through the integrated logic circuits in the hardware of the processor element or through software instructions.

[0221] Figure 12 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 12 As shown, the electronic device may include: transceiver 121, processor 122, and memory 123.

[0222] Processor 122 executes computer execution instructions stored in memory, causing processor 122 to perform the scheme in the above embodiments. Processor 122 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0223] The memory 123 is connected to the processor 122 via the system bus and completes communication between them. The memory 123 is used to store computer program instructions.

[0224] Transceiver 121 can be used to obtain the task to be run and its configuration information.

[0225] The system bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0226] The electronic device provided in this application embodiment can be the information recommendation device of the above embodiment, or it can be the training device for the research report information scoring prediction model of the above embodiment.

[0227] This application also provides a chip for executing instructions, which is used to execute the technical solution of the information recommendation method in the above embodiments, or to execute the technical solution of the training method of the research report information scoring prediction model in the above embodiments.

[0228] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, they cause the computer to perform the technical solution of the information recommendation method described in the above embodiments, or to perform the technical solution of the training method of the research report information scoring prediction model described in the above embodiments.

[0229] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the information recommendation method in the above embodiments, or implement the technical solution of the training method of the research report information scoring prediction model in the above embodiments.

[0230] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0231] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. An information recommendation method, characterized in that, include: Obtain the behavior sequence of the target object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the target object in response to the research report information; Each session information is encoded to obtain a first feature vector for each session information; Extract the temporal distance feature between behaviors from multiple first feature vectors of the multiple session information to obtain a second feature vector; Based on the multi-head self-attention mechanism, the spatial distance feature of the true intent is extracted from the second feature vector to obtain the preference vector of the target object; Based on the preference vector, a research report recommendation list for the target object is obtained from at least one candidate research report information, and the research report recommendation list is recommended to the target object. The step of extracting temporal distance features between behaviors from multiple first feature vectors of the multiple session information to obtain a second feature vector includes: Based on a preset independent time interval gated loop unit, the time distance between behaviors is extracted from multiple first feature vectors of the multiple session information to obtain the second feature vector; The independent time interval gated loop unit is a gated loop unit structure that incorporates a time gate to separately store time interval information.

2. The method as described in claim 1, characterized in that, The calculation process of the independent time interval gated loop unit is as follows: in, It is a sigmoid function. The time interval between the action sequence and the current time. The preset vector activation parameters, The state is hidden for the time interval of the current action sequence. To reset the door, To update the door, For the preset parameter matrix, The candidate hidden state for the current action sequence time interval. The OR operator; ,and , This indicates a time gate, which represents the effect of a time interval on the independent time interval gated loop unit; where... and Simultaneously controlling the hidden state of the current action sequence. And assign a greater weight to historical behaviors that are close to the target.

3. The method as described in claim 1, characterized in that, The step of extracting spatial distance features of the true intent from the second feature vector based on a multi-head self-attention mechanism to obtain the preference vector of the target object includes: Based on the multi-head self-attention mechanism and the second feature vector, the self-attention value output by each attention head is obtained; The weight ratio of each attention head is obtained based on the self-attention value output by each attention head and the second feature vector; The preference vector of the target object is obtained based on the self-attention value of each attention head and the weight ratio of each attention head.

4. The method as described in claim 3, characterized in that, The formula for calculating the weight percentage of each attention head is as follows: in, For the first The weighting of each attention head, The total number of attention heads. For the first The transpose of the self-attention value output by the attention head. This is the second feature vector.

5. The method as described in claim 3 or 4, characterized in that, The step of obtaining the preference vector of the target object based on the self-attention value of each attention head and the weight ratio of each attention head includes: The preference vector of the target object is obtained by weighting and summing the self-attention value of each attention head with its corresponding weight ratio.

6. The method as described in claim 1, characterized in that, The step of obtaining a recommendation list of research reports for the target object from at least one candidate research report based on the preference vector includes: Obtain at least one candidate research report and a vector for each candidate research report; Based on the preference vector and the vector of each candidate research report, obtain the predicted score value for each candidate research report; Based on the predicted score of each candidate research report, a recommended list of research reports for the target object is obtained from the at least one candidate research report.

7. The method as described in claim 6, characterized in that, The formula for calculating the predicted score of each candidate research report is as follows: in, For the first The predicted score values ​​for each of the candidate research reports. For the preference vector, For the first A vector of the candidate research report information. For behavioral bias terms, For candidate research report information bias items, This is a pre-trained weight matrix.

8. A training method for a research report information scoring prediction model, characterized in that, The research report information rating prediction model is applicable to the information recommendation method as described in any one of claims 1 to 7, and the training method includes: Obtain the behavior sequence of the sample object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the sample object in response to the research report information; The multiple conversation information is input into the initial research report information rating prediction model; wherein, the research report information rating prediction model includes an encoding module, a temporal distance feature extraction module, a spatial distance feature extraction module for true intent, and a prediction module; The encoding module encodes each session information to obtain a first feature vector for each session information. Based on the time distance feature extraction module, the time distance feature between behaviors is extracted from the multiple first feature vectors of the multiple session information to obtain the second feature vector; Based on the spatial distance feature extraction module of the true intent, a multi-head self-attention mechanism is used to extract the spatial distance feature of the true intent from the second feature vector to obtain the preference vector of the sample object. The preference vector is sent to the prediction module to obtain the predicted score value for each candidate research report, and the research report score prediction model is trained based on the predicted score value for each candidate research report.

9. The method as described in claim 8, characterized in that, The time distance feature extraction module includes an independent time interval gated loop unit; wherein, the calculation process of the independent time interval gated loop unit is as follows: in, It is a sigmoid function. The time interval between the action sequence and the current time. The preset vector activation parameters, The state is hidden for the time interval of the current action sequence. To reset the door, To update the door, For the preset parameter matrix, The candidate hidden state for the current action sequence time interval. The OR operator; ,and , This indicates a time gate, which represents the effect of a time interval on the independent time interval gated loop unit; where... and Simultaneously controlling the hidden state of the current action sequence. And assign a greater weight to historical behaviors that are close to the target.

10. The method as described in claim 8, characterized in that, The spatial distance feature extraction module based on the true intent employs a multi-head self-attention mechanism to extract the spatial distance features of the true intent from the second feature vector, thereby obtaining the preference vector of the sample object, including: The spatial distance feature extraction module of the true intent obtains the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector, and obtains the weight ratio of each attention head according to the self-attention value output by each attention head and the second feature vector, and obtains the preference vector of the target object according to the self-attention value of each attention head and the weight ratio of each attention head.

11. The method as described in claim 8, characterized in that, The formula for the prediction module is expressed as follows: in, The first output of the prediction module The predicted score values ​​for each of the candidate research reports. For the preference vector, For the first A vector of the candidate research report information. For behavioral bias terms, For candidate research report information bias items, This is the weight matrix to be trained.

12. The method as described in claim 8, characterized in that, The step of training the research report information scoring prediction model based on the predicted scoring value of each candidate research report information includes: A preset cross-loss function is used to calculate the cross-loss function of the research report information score prediction model based on the score prediction value of each candidate research report information; The cross-loss function is used as the objective function, and the stochastic gradient descent algorithm is used to optimize the objective function. The error backpropagation algorithm is then used to update the model parameters of the research report information scoring prediction model.

13. An information recommendation device, characterized in that, include: The processing module is used to acquire the behavior sequence of the target object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the target object in response to the research report information; An encoding module is used to encode each of the session information to obtain a first feature vector for each of the session information; The first extraction module is used to extract the temporal distance feature between behaviors from multiple first feature vectors of the multiple session information to obtain a second feature vector. The second extraction module is used to extract the spatial distance features of the true intent from the second feature vector based on a multi-head self-attention mechanism, so as to obtain the preference vector of the target object. The recommendation module is used to obtain a research report recommendation list of the target object from at least one candidate research report information according to the preference vector, and recommend the research report recommendation list to the target object; The first extraction module is specifically used for: Based on a preset independent time interval gated loop unit, the time distance between behaviors is extracted from multiple first feature vectors of the multiple session information to obtain the second feature vector; The independent time interval gated loop unit is a gated loop unit structure that incorporates a time gate to separately store time interval information.

14. The apparatus as claimed in claim 13, characterized in that, The calculation process of the independent time interval gated loop unit is as follows: in, It is a sigmoid function. The time interval between the action sequence and the current time. The preset vector activation parameters, The state is hidden for the time interval of the current action sequence. To reset the door, To update the door, For the preset parameter matrix, The candidate hidden state for the current action sequence time interval. The OR operator; ,and , This indicates a time gate, which represents the effect of a time interval on the independent time interval gated loop unit; where... and Simultaneously controlling the hidden state of the current action sequence. And assign a greater weight to historical behaviors that are close to the target.

15. The apparatus as claimed in claim 13, characterized in that, The second extraction module is specifically used for: Based on the multi-head self-attention mechanism and the second feature vector, the self-attention value output by each attention head is obtained; The weight ratio of each attention head is obtained based on the self-attention value output by each attention head and the second feature vector; The preference vector of the target object is obtained based on the self-attention value of each attention head and the weight ratio of each attention head.

16. The apparatus as claimed in claim 15, characterized in that, The formula for calculating the weight percentage of each attention head is as follows: in, For the first The weighting of each attention head, The total number of attention heads. For the first The transpose of the self-attention value output by the attention head. This is the second feature vector.

17. The apparatus as claimed in claim 15 or 16, characterized in that, The second extraction module is specifically used for: The preference vector of the target object is obtained by weighting and summing the self-attention value of each attention head with its corresponding weight ratio.

18. The apparatus as claimed in claim 13, characterized in that, The recommendation module is specifically used for: Obtain at least one candidate research report and a vector for each candidate research report; Based on the preference vector and the vector of each candidate research report, obtain the predicted score value for each candidate research report; Based on the predicted score of each candidate research report, a recommended list of research reports for the target object is obtained from the at least one candidate research report.

19. The apparatus as claimed in claim 18, characterized in that, The formula for calculating the predicted score of each candidate research report is as follows: in, For the first The predicted score values ​​for each of the candidate research reports. For the preference vector, For the first A vector of the candidate research report information. For behavioral bias terms, For candidate research report information bias items, This is a pre-trained weight matrix.

20. A training device for a research report information scoring prediction model, characterized in that, The research report information rating prediction model is applicable to the information recommendation method as described in any one of claims 1 to 7, and the training device includes: The processing module is used to acquire the behavior sequence of the sample object and divide the behavior sequence into multiple session information; wherein, the behavior sequence includes at least one behavior information, and each behavior information is the behavior data of the sample object in response to the research report information; An input module is used to input the multiple session information into an initial research report information rating prediction model; wherein, the research report information rating prediction model includes an encoding module, a temporal distance feature extraction module, a spatial distance feature extraction module for true intent, and a prediction module; An encoding module is used to encode each session information based on the encoding module to obtain a first feature vector for each session information; The first extraction module is used to extract time distance features between behaviors from multiple first feature vectors of the multiple session information based on the time distance feature extraction module, so as to obtain a second feature vector. The second extraction module is used to extract the spatial distance features of the true intent based on the spatial distance feature extraction module of the true intent by using a multi-head self-attention mechanism to extract the spatial distance features of the true intent from the second feature vector in order to obtain the preference vector of the sample object. The training module is used to send the preference vector to the prediction module to obtain the predicted score value of each candidate research report information, and to train the research report information score prediction model based on the predicted score value of each candidate research report information.

21. The apparatus as claimed in claim 20, characterized in that, The time distance feature extraction module includes an independent time interval gated loop unit; wherein, the calculation process of the independent time interval gated loop unit is as follows: in, It is a sigmoid function. The time interval between the action sequence and the current time. The preset vector activation parameters, The state is hidden for the time interval of the current action sequence. To reset the door, To update the door, For the preset parameter matrix, The candidate hidden state for the current action sequence time interval. The OR operator; ,and , This indicates a time gate, which represents the effect of a time interval on the independent time interval gated loop unit; where... and Simultaneously controlling the hidden state of the current action sequence. And assign a greater weight to historical behaviors that are close to the target.

22. The apparatus as claimed in claim 20, characterized in that, The second extraction module is specifically used for: The spatial distance feature extraction module of the true intent obtains the self-attention value output by each attention head based on the multi-head self-attention mechanism and the second feature vector, and obtains the weight ratio of each attention head according to the self-attention value output by each attention head and the second feature vector, and obtains the preference vector of the target object according to the self-attention value of each attention head and the weight ratio of each attention head.

23. The apparatus as claimed in claim 20, characterized in that, The formula for the prediction module is expressed as follows: in, The first output of the prediction module The predicted score values ​​for each of the candidate research reports. For the preference vector, For the first A vector of the candidate research report information. For behavioral bias terms, For candidate research report information bias items, This is the weight matrix to be trained.

24. The apparatus as claimed in claim 20, characterized in that, The step of training the research report information scoring prediction model based on the predicted scoring value of each candidate research report information includes: A preset cross-loss function is used to calculate the cross-loss function of the research report information score prediction model based on the score prediction value of each candidate research report information; The cross-loss function is used as the objective function, and the stochastic gradient descent algorithm is used to optimize the objective function. The error backpropagation algorithm is then used to update the model parameters of the research report information scoring prediction model.

25. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7, or to implement the method as described in any one of claims 8 to 12.

26. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7, or to implement the method as described in any one of claims 8 to 12.

27. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7, or implements the steps of the method according to any one of claims 8 to 12.

Images