Decision style driven decision scheme recommendation method

Abstract

The invention discloses a decision-making scheme recommendation method driven by a decision-making style, which relates to the technical field of intelligent recommendation, and comprises the following steps of: acquiring feature vectors of attributes in a decision-making scheme after comprehensively considering all attribute information by adopting a style-driven scheme evaluation network based on an attention mechanism; the feature vectors are obtained after attention mechanism fusion and all attribute information is comprehensively considered, the probability that a decision scheme belongs to a decision style is obtained, and then a trained style-driven scheme evaluation network is obtained; and finally, obtaining a score of a new task decision scheme according to the trained style-driven scheme evaluation network, and realizing recommendation of the new task decision scheme. According to the method, the scheme according with the decision-making style of the user is adopted to evaluate the network for recommendation according to the decision-making style of the individual user, the recommended decision-making scheme can be consistent with the decision-making preference style of the individual user to the maximum extent, and the consistency between the recommendation result and the individual user is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to a decision-style-driven method for recommending decision-making solutions. Background Technology

[0002] To address the issues of rigid human-computer interaction, closed machine computation processes, and lack of interpretability in current task-assisted decision-making systems, this paper proposes a human-in-the-loop (HIL) HIL system that supports adaptive, mutually trusting, and collaborative decision-making in maritime environments. This system aims to foster human-computer collaborative decision-making capabilities characterized by "style accompaniment + enhanced computing power." When selecting from multiple decision options for a task, existing solution evaluation networks primarily recommend solutions based on the inherent attributes of the options themselves. However, when making recommendations to individual users, the chosen solutions fail to reflect their individual decision-making preferences and styles. Summary of the Invention

[0003] This invention discloses a microservice recommendation method based on a decision style-driven strategy to overcome the aforementioned technical problems.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A decision-style-driven method for recommending decision-making solutions includes the following steps:

[0006] S1: Obtain multiple decision options for the user's historical tasks to obtain the attributes of the decision options;

[0007] S2: Based on the attributes of the decision scheme, obtain the feature vectors of the attributes of the decision scheme, including: the latent vector representation corresponding to the attention query matrix of the attributes of the decision scheme, the latent vector representation corresponding to the attention key matrix, and the latent vector representation corresponding to the attention value matrix.

[0008] S3: Based on the latent vector representations of the attention query matrix and the attention key matrix of the decision scheme's attributes, obtain the similarity between the j-th attribute and the i-th attribute of the decision scheme; where i and j are the index numbers of the attributes of the decision scheme; I represents the total number of attributes of the decision scheme;

[0009] S4: Based on the similarity between the j-th attribute and the i-th attribute of the decision scheme, obtain the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme;

[0010] S5: Based on the similarity between the j-th attribute and the ith attribute of the normalized decision scheme and the latent vector representation corresponding to the attention value matrix, obtain the feature vector of the j-th attribute in the decision scheme after comprehensively considering all attribute information, so as to obtain the feature vector obtained after comprehensively considering all attribute information after fusion by the Self-attention mechanism.

[0011] S6: Based on the feature vector obtained after fusion through the Self-attention mechanism and after comprehensively considering all attribute information, obtain the probability that the decision scheme belongs to the m-th style based on the Gaussian mixture model; determine the style to which the decision scheme belongs, and train the style-driven scheme evaluation network; m is the index number of the decision style;

[0012] S7: Obtain the user's new task decision scheme, evaluate the network based on the trained style-driven scheme, obtain the feature vector of the attributes in the new task decision scheme after comprehensively considering all attribute information; then obtain the score of the new task decision scheme; and make recommendations based on the score of the new task decision scheme.

[0013] Furthermore, the formula used to score the new task decision-making scheme is as follows:

[0014]

[0015] In the formula: S subj The score for the decision-making options for the new task; γ i′ It is the attention weight of the feature vector obtained by comprehensively considering all attribute information for the i′th attribute in the new decision scheme; w i′ For b i′ The weighting parameters; b i′ This represents the feature vector of the i′-th attribute in the new decision scheme after comprehensively considering all attribute information; is the bias term; tanh represents the hyperbolic tangent function; i′ is the index number of the attribute of the new task; I′ is the total number of attributes of the new task.

[0016] Furthermore, the formula used to obtain the feature vectors of the attributes of the decision-making scheme is as follows:

[0017] q i =W q ·x i , i = 1, ... I

[0018] k i =W k ·x i , i = 1, ... I

[0019] vi =W v ·x i , i = 1, ... I

[0020] In the formula: x i q represents the i-th attribute of the decision-making scheme; i The latent vector representation of the attention query matrix corresponding to the i-th attribute of the decision scheme; k i The latent vector representation of the attention key matrix corresponding to the i-th attribute of the decision scheme; v i W represents the latent vector representation of the attention value matrix corresponding to the i-th attribute of the decision scheme. q W represents the query matrix weights in the attention mechanism. k W represents the weights of the key matrix in the attention mechanism. v This represents the weights of the value matrix in the attention mechanism.

[0021] Furthermore, the formula used to obtain the similarity between the j-th attribute and the i-th attribute of the decision scheme is as follows:

[0022]

[0023] In the formula: a j，i q represents the similarity between the j-th attribute and the i-th attribute of the decision-making scheme; j The latent vector representation of the attention query matrix corresponding to the j-th attribute of the decision scheme; k i Let represent the latent vector representation corresponding to the attention key matrix of the i-th attribute of the decision scheme; D represents the total number of dimensions of the latent vectors obtained after matrix transformation; d represents the index number of the dimension of the latent vectors obtained after matrix transformation; q j，d k represents the value of the d-th dimension component in the latent vector corresponding to the attention query matrix of the j-th attribute of the decision scheme; i，d The value of the d-th dimension component in the latent vector corresponding to the attention key matrix of the i-th attribute of the decision scheme is represented; i and j both represent the index number of the attribute of the decision scheme; I represents the total number of attributes of the decision scheme.

[0024] Furthermore, the formula used to obtain the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme is as follows:

[0025]

[0026] In the formula: represents the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; exp(·) represents the exponential function.

[0027] Furthermore, the formula used to obtain the feature vector after the fusion through the Self-attention mechanism, which takes into account all attribute information, is as follows;

[0028]

[0029]

[0030] In the formula: b j This represents the feature vector of the j-th attribute in the decision-making scheme after comprehensively considering all attribute information; This represents the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; v i B represents the latent vector representation corresponding to the attention value matrix of the i-th attribute of the decision scheme; B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information.

[0031] Furthermore, the formula used to obtain the probability that the decision option belongs to the m-th style is as follows:

[0032]

[0033] In the formula: M is the total number of decision styles; m is the index number of the decision style; B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information; π m Let m be the prior probability of the m-th decision style category; Indicates the mean μ m With covariance ∑ m Let B be a Gaussian probability density distribution; P(B) represents the probability that the decision scheme belongs to the m-th style.

[0034] Furthermore, the attributes of the decision-making scheme include: task category, route, task objective, number of task execution entities, task execution tools used, and decision time.

[0035] Beneficial Effects: This invention provides a decision-making style-driven solution recommendation method. It employs a style-driven solution evaluation network based on an attention mechanism. By acquiring the similarity between attributes of decision-making solutions, it obtains the feature vectors of attributes in each solution after comprehensively considering all attribute information. This feature vector is then obtained after attention-based fusion and comprehensive consideration of all attribute information. Furthermore, based on a Gaussian mixture model, the probability of a decision-making solution belonging to a particular decision-making style is obtained, leading to the trained style-driven solution evaluation network. Finally, the trained style-driven solution evaluation network is used to obtain a score for a new task decision-making solution, thus achieving the recommendation of a decision-making solution for a new task. This invention trains the solution evaluation network by considering the user's style to which the decision-making solution belongs. This allows the predicted results of the trained style-driven solution evaluation network to be tailored to the individual user's decision-making style, recommending solutions that align with that style. The recommended decision-making solutions are maximally consistent with the individual user's decision-making preferences, improving the consistency between the recommendation results and the individual user. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 The flowchart of the decision-making scheme recommendation method of the present invention is shown below;

[0038] Figure 2 This is a schematic diagram illustrating the feature vector calculation process for obtaining decision scheme attributes based on the attention mechanism in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram illustrating the principle of parallel computing in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the style-driven scheme evaluation network structure in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This embodiment introduces a decision-style-driven method for recommending decision-making solutions, such as... Figure 1 As shown, it includes the following steps:

[0043] S1: Obtain multiple decision options for the user's historical tasks to obtain the attributes of the decision options;

[0044] Specifically, the attributes of the decision-making scheme in this embodiment include: task category, route, task objective, number of task execution entities, task execution tools used, decision time, etc.; and then, based on the attention encoder, the feature vector of the user's decision-making scheme is obtained.

[0045] Specifically, the method used to obtain the attributes of multiple decision-making schemes is prior art to those skilled in the art and will not be described in detail here. The attributes of the decision-making schemes, including task category, route, task objective, number of task executors, task execution tools used, and decision time, are respectively labeled as x1, x2, x3, ..., x... I Represented. Then using matrix W q , W k, W v respectively with x1, x2, x2, ..., x I Multiplication, x I Let q represent the i-th attribute of the decision-making scheme. i k i v i ,i∈(1,2,3,...,I), such as Figure 2 As shown.

[0046] Specifically, in this embodiment, each attribute of the decision scheme is extracted, and each attribute is embedded and encoded using three methods: label encoding, ordinal encoding, and hybrid encoding, as input to the style-driven scheme evaluation network.

[0047] S2: Based on the attributes of the decision scheme, the attention encoder in the style-driven scheme evaluation network based on the attention mechanism is used to obtain the feature vectors of the attributes of the decision scheme, including: the latent vector representations corresponding to the attention query matrix of the attributes of the decision scheme, the latent vector representations corresponding to the attention key matrix, and the latent vector representations corresponding to the attention value matrix.

[0048] Preferably, the formula used to obtain the feature vector of the attributes of the decision scheme is as follows:

[0049]

[0050] In the formula: x i Let q represent the i-th attribute of the decision-making scheme (such as task category, route, task objective, etc.);i The latent vector representation of the attention query matrix corresponding to the i-th attribute of the decision scheme; k i The latent vector representation of the attention key matrix corresponding to the i-th attribute of the decision scheme; v i W represents the latent vector representation of the attention value matrix corresponding to the i-th attribute of the decision scheme. q W represents the query matrix weights in the attention mechanism. k W represents the weights of the key matrix in the attention mechanism. v This represents the weights of the value matrix in the attention mechanism.

[0051] S3: The latent vector representation q corresponding to the attention query matrix of the i-th, i=1,…I-th attribute of the decision scheme. j The latent vector representation k corresponding to the attention key matrix of the i-th, i=1,…I attribute i Obtain the similarity α between the j-th attribute and the i-th attribute of the decision-making scheme. j，i ;

[0052]

[0053] In the formula: a j，i q represents the similarity between the j-th attribute and the i-th attribute of the decision-making scheme; j The latent vector representation of the attention query matrix corresponding to the j-th attribute of the decision scheme; k i Let represent the latent vector representation corresponding to the attention key matrix of the i-th attribute of the decision scheme; D represents the total number of dimensions of the latent vectors obtained after matrix transformation; d represents the index number of the dimension of the latent vectors obtained after matrix transformation; q j,d k represents the value of the d-th dimension component in the latent vector corresponding to the attention query matrix of the j-th attribute of the decision scheme; i,d The value of the d-th dimension component in the latent vector corresponding to the attention key matrix of the i-th attribute of the decision scheme is represented; i and j both represent the index number of the attribute of the decision scheme; I represents the total number of attributes of the decision scheme.

[0054] S4: Based on the similarity a between the j-th attribute and the i-th attribute of the decision scheme. j,i Obtain the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme.

[0055] Specifically, the similarity α between the j-th attribute and the i-th attribute of the decision-making scheme is... j，i Input softmax layer, obtain Based on this method, the normalized similarity between other attributes of the decision-making schemes is obtained;

[0056]

[0057] in, represents the similarity between the j-th attribute and the ith attribute of the normalized decision scheme, i.e., the normalized similarity weight between the 1st attribute and the ith attribute; exp(·) represents the exponential function;

[0058] S5: Based on the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme. The latent vector representation corresponding to the attention value matrix is ​​used to obtain the feature vector b of the j-th attribute in the decision scheme after comprehensively considering all attribute information. j To obtain the feature vector obtained after the fusion through the Self-attention mechanism, which takes into account all attribute information;

[0059]

[0060] In the formula: b j This represents the feature vector of the j-th attribute in the decision-making scheme after comprehensively considering all attribute information; This represents the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; v i B represents the latent vector representation corresponding to the attention value matrix of the i-th attribute of the decision scheme; B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information.

[0061] Specifically, the result obtained in the previous step Corresponding to the corresponding positions v1, v2, v3, ..., v I Multiply and then sum to obtain the output b1 corresponding to the first attribute x1 in the input decision scheme. Based on this, we can obtain the feature vector B of all decision schemes obtained after the fusion through the Self-attention mechanism, which comprehensively considers all attribute information.

[0062] Specifically, in RNNs or LSTMs, the input sequence is input sequentially step-by-step, meaning that the computation at each time step depends on the result of the computation at the previous time step. This sequential processing makes it impossible to fully parallelize on hardware, resulting in slowness when processing long sequences or large-scale data. For the input x1, x2, x3, ..., x... I Unlike the processing of RNN / LSTM, this embodiment uses a self-attention mechanism, which can process x1, x2, x3, ..., x in parallel. I This greatly improves the computational efficiency for x1, x2, x3, ..., xI The speed of feature extraction is improved, making it more efficient than RNN / LSTM in feature extraction. Combining the above Self-attention calculation process, the principle of parallel computation is as follows: Figure 3 As shown.

[0063] Specifically, in this embodiment, each attribute (e.g., task category, route, task objective, etc.) in the decision-making scheme is represented as an input vector sequence x1, x2, x3, ..., x I The input matrix X is formed by multiplying it by matrix W. q W k W v Transforming the input matrix X yields three matrices Q, K, and W:

[0064] Q = X·W q K = X·W k V = X·W v

[0065] Specifically, by calculating the dot product of the query matrix Q and the key matrix K, the similarity score matrix between each attribute can be obtained, and a scaling factor can be used to... (where d is the dimension of the mapped latent vector), and the attention score matrix is ​​calculated.

[0066]

[0067] and

[0068]

[0069] Next, the similarity score of each row is normalized using the softmax function to obtain the normalized attention weight matrix:

[0070]

[0071] Finally, the normalized attention weight matrix is ​​multiplied by the value matrix V to obtain the final scheme feature vector B:

[0072]

[0073] In each row B i It gathers the weighted feature information of all attributes, reflecting the final result of the information interaction between different attributes of the original input.

[0074] S6: Based on the feature vector obtained after fusion through the Self-attention mechanism and after comprehensively considering all attribute information, the probability P(B) of the decision scheme belonging to the m-th style is obtained based on the Gaussian mixture model; to determine the style to which the n-th decision scheme belongs; and thus to obtain the trained style-driven scheme evaluation network.

[0075] In this embodiment, a Gaussian mixture model based on decision style is used to divide the feature vector B obtained after fusion through the self-attention mechanism and comprehensive consideration of all attribute information into three categories, representing the radical, balanced, and conservative styles to which the decision scheme belongs, respectively. The Gaussian mixture model based on decision style is existing technology in the field, and this embodiment only uses it to achieve the desired functionality.

[0076]

[0077] In the formula: M represents the total number of decision styles, corresponding to Gaussian distributions of aggressive, balanced, and conservative solutions; m is the index number of the decision style (1 represents aggressive, 2 represents balanced, and 3 represents conservative); B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information; π m Let m be the prior probability of the m-th decision style category; Indicates the mean μ m With covariance ∑ m Let B be a Gaussian probability density distribution; P(B) represents the probability that the decision scheme belongs to the m-th style.

[0078] Specifically, the structure of the style-driven scheme evaluation network in this embodiment is as follows: Figure 4 As shown, it includes: an Embedding layer that embeds and encodes the scheme, using methods such as label encoding, sequence encoding, and hybrid encoding; an Encoder layer that uses an attention mechanism to extract feature information, amplifying the weights of important features to prepare for subsequent scheme evaluation; and an Evaluate layer that integrates feature information through several omnidirectional connected layers and outputs a scheme score to evaluate the scheme.

[0079] Specifically, after obtaining the probability P(B) that the decision scheme belongs to the m-th style, the accuracy of the prediction result of the style-driven scheme evaluation network can be determined based on the true style of the decision scheme. Only the network parameters of the style-driven scheme evaluation network can be adjusted to obtain the network parameters of the trained style-driven scheme evaluation network. This is a conventional technique for those skilled in the art, and therefore will not be described in detail.

[0080] S7: Obtain the user's new task decision plan, and obtain the feature vector of the attributes in the new task decision plan after comprehensively considering all attribute information; then obtain the score of the new task decision plan; and make recommendations based on the score of the new task decision plan.

[0081] The formula used is as follows:

[0082]

[0083] In the formula: S subj The score for the decision-making options for the new task; γ i′ It is the attention weight of the feature vector obtained by comprehensively considering all attribute information for the i′th attribute in the new decision scheme; w i′ For b i′ The weighting parameters; b i′ This represents the feature vector of the i′-th attribute in the new decision scheme after comprehensively considering all attribute information; is the bias term; tanh represents the hyperbolic tangent function; i′ is the index number of the attribute of the new task; I′ is the total number of attributes of the new task.

[0084] Specifically, because users develop preferences for certain decision-making options due to subjective reasons, these preferences are not only related to the current state but also influenced by the user's decision-making intent. This embodiment categorizes the user's decision-making intent into aggressive, balanced, and conservative styles based on their personal strategy preferences. By learning the user's style through historical interactions with the solution set, this embodiment transforms these three styles into evaluations of decision-making options using a style-driven solution evaluation network. This simulates the user's subjective preferences and addresses the ambiguity in the definitions of different styles. Specifically, aggressive, balanced, and conservative styles are artificially defined style types, and classifying styles is essentially a clustering process, which will lose some precise user characteristics. For some boundary decision-making behaviors (i.e., behaviors between two styles), the accuracy of this discretized evaluation result will be significantly reduced. This embodiment uses a style-driven solution evaluation network to convert the discretized output into continuous numerical values, enabling the accurate extraction of the user's personal strategy preferences.

[0085] Specifically, even with the same style, users' specific decision-making behaviors can differ. Using style as the training objective can easily lead to training instability and prevent recommendations from being fully based on user preferences during actual implementation. In this embodiment, decision schemes are evaluated and scores are output. This is a bijective process, meaning that for each score, there is one and only one corresponding scheme, theoretically resulting in better training performance. Therefore, this embodiment can transform decision style and task attributes into scheme evaluation through a unified network structure. For example, for a specific decision scheme, if the user's style matches the style of the decision scheme, the scheme's score is relatively high; if the styles do not match, the decision scheme's score is relatively low. Therefore, the problem is transformed into: training a style-driven scheme evaluation network (EvalNet) to score decision schemes and finding the highest-scoring decision scheme as the one that best matches the user's personal strategy preferences, where the network parameters represent the user's preference parameters. The structure of the style-driven scheme evaluation network EvalNet is as follows: Figure 4 As shown.

[0086] Specifically, the Embedding layer in the style-driven solution evaluation network embeds and encodes the solutions using methods such as label encoding, ordinal encoding, and hybrid encoding. The Encoder layer uses an attention mechanism to extract feature information and amplify the weights of important features, laying the groundwork for subsequent solution evaluation. The Evaluate layer integrates feature information through several omnidirectional connected layers and outputs a solution score to evaluate the solution.

[0087] Specifically, a style-driven solution evaluation network is established for each user. If there is no user history information, a user's initial style is inferred through a questionnaire, and a style-driven solution evaluation network with a typical style is assigned to that user. During human-computer interaction, the style-driven solution evaluation network calculates the style of each candidate solution, and the user preference network infers the user's preference coefficient, i.e., network parameters, based on the decision solution and its corresponding style. Then, the style-driven solution evaluation network based on that user is updated based on user feedback, thereby realizing a decision preference inverse learning technique that incorporates user feedback.

[0088] Specifically, when a user enters the system, the system matches the user with their corresponding account. If the match is successful, the system directly calls the user's style-driven solution evaluation network. This style-driven solution evaluation network is the network calculated and stored locally in a file during the user's previous interactions with the system. If the match fails, the system preliminarily determines the user's style through a questionnaire, matches a typical style-driven solution evaluation network corresponding to that style, and initializes it. The user ID or account is stored locally as a filename for matching purposes.

[0089] Specifically, the system needs to recommend solutions that match the user's style; on the other hand, the system also needs to guide the user to make the most advantageous decision. Therefore, this embodiment calculates the scores of multiple decision solutions for a decision task, sorts the decision solutions from highest to lowest score, and recommends them to the user, awaiting the user's decision. This completes the intelligent sorting process for decision solutions to a new task and the recommendation of decision solutions.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A decision-style-driven method for recommending decision-making solutions, characterized in that, Includes the following steps: S1: Obtain multiple decision options for the user's historical tasks to obtain the attributes of the decision options; S2: Based on the attributes of the decision scheme, obtain the feature vectors of the attributes of the decision scheme, including: the latent vector representation corresponding to the attention query matrix of the attributes of the decision scheme, the latent vector representation corresponding to the attention key matrix, and the latent vector representation corresponding to the attention value matrix. S3: Based on the latent vector representations of the attention query matrix and the attention key matrix of the decision scheme's attributes, obtain the similarity between the j-th attribute and the i-th attribute of the decision scheme; where i and j are the index numbers of the attributes of the decision scheme; I represents the total number of attributes of the decision scheme; S4: Based on the similarity between the j-th attribute and the i-th attribute of the decision scheme, obtain the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; S5: Based on the similarity between the j-th attribute and the ith attribute of the normalized decision scheme and the latent vector representation corresponding to the attention value matrix, obtain the feature vector of the j-th attribute in the decision scheme after comprehensively considering all attribute information, so as to obtain the feature vector obtained after comprehensively considering all attribute information after fusion by the Self-attention mechanism. S6: Based on the feature vector obtained after fusion through the Self-attention mechanism and after comprehensively considering all attribute information, obtain the probability that the decision scheme belongs to the m-th style based on the Gaussian mixture model; determine the style to which the decision scheme belongs, and train the style-driven scheme evaluation network; m is the index number of the decision style; S7: Obtain the user's new task decision scheme, evaluate the network based on the trained style-driven scheme, obtain the feature vector of the attributes in the new task decision scheme after comprehensively considering all attribute information; then obtain the score of the new task decision scheme; and make recommendations based on the score of the new task decision scheme.

2. The decision-style-driven decision recommendation method according to claim 1, characterized in that, The formula used to score new task decision options is as follows: In the formula: S subj The score for the decision-making options for the new task; γ i′ It is the attention weight of the feature vector obtained by comprehensively considering all attribute information for the i′th attribute in the new decision scheme; w i′ For b i′ The weighting parameters; b i′ This represents the feature vector of the i′-th attribute in the new decision scheme after comprehensively considering all attribute information; is the bias term; tanh represents the hyperbolic tangent function; i′ is the index number of the attribute of the new task; I′ is the total number of attributes of the new task.

3. The decision-style-driven decision recommendation method according to claim 1, characterized in that, The formula used to obtain the feature vectors of the attributes of the decision-making scheme is as follows: q i =W q ·x i ,i=1,…I k i =W k ·x i ,i=1,…I v i =W v ·x i ,i=1,…I In the formula: x i q represents the i-th attribute of the decision-making scheme; i The latent vector representation of the attention query matrix corresponding to the i-th attribute of the decision scheme; k i The latent vector representation of the attention key matrix corresponding to the i-th attribute of the decision scheme; v i W represents the latent vector representation of the attention value matrix corresponding to the i-th attribute of the decision scheme. q This represents the query matrix weights in the attention mechanism; W k W represents the weights of the key matrix in the attention mechanism. v This represents the weights of the value matrix in the attention mechanism.

4. The decision-style-driven decision recommendation method according to claim 1, characterized in that, The formula used to obtain the similarity between the j-th attribute and the i-th attribute of the decision-making scheme is as follows: In the formula: a j，i q represents the similarity between the j-th attribute and the i-th attribute of the decision-making scheme; j The latent vector representation of the attention query matrix corresponding to the j-th attribute of the decision scheme; k i Let represent the latent vector representation corresponding to the attention key matrix of the i-th attribute of the decision scheme; D represents the total number of dimensions of the latent vectors obtained after matrix transformation; d represents the index number of the dimension of the latent vectors obtained after matrix transformation; q j，d k represents the value of the d-th dimension component in the latent vector corresponding to the attention query matrix of the j-th attribute of the decision scheme; i，d The value of the d-th dimension component in the latent vector corresponding to the attention key matrix of the i-th attribute of the decision scheme is represented; i and j both represent the index number of the attribute of the decision scheme; I represents the total number of attributes of the decision scheme.

5. The decision-style-driven decision recommendation method according to claim 4, characterized in that, The formula used to obtain the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme is as follows: In the formula: represents the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; exp(·) represents the exponential function.

6. The decision-style-driven decision recommendation method according to claim 5, characterized in that, The formula used to obtain the feature vector after fusion through the Self-attention mechanism, which takes into account all attribute information, is as follows; In the formula: b j This represents the feature vector of the j-th attribute in the decision-making scheme after comprehensively considering all attribute information; This represents the similarity between the j-th attribute and the i-th attribute of the normalized decision scheme; v i B represents the latent vector representation corresponding to the attention value matrix of the i-th attribute of the decision scheme; B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information.

7. The decision-style-driven decision recommendation method according to claim 5, characterized in that, The formula used to obtain the probability that a decision belongs to the m-th style is as follows: In the formula: M is the total number of decision styles; m is the index number of the decision style; B represents the feature vector obtained after fusion through the Self-attention mechanism, taking into account all attribute information; π m Let be the prior probability of the m-th decision style category; Indicates the mean μ m With covariance ∑ m Let B be a Gaussian probability density distribution; P(B) represents the probability that the decision scheme belongs to the m-th style.

8. The decision-style-driven decision recommendation method according to claim 1, characterized in that, The attributes of the decision-making scheme include: task category, route, task objective, number of task execution entities, task execution tools used, and decision time.

Images