Surprise recommendation method and device, storage medium and computer program product
By generating explanatory text through a multi-task learning framework and user comment text, the problems of insufficient accuracy and explainability in existing surprise recommendations are solved, and a more accurate and explainable recommendation effect is achieved.
Patent Information
- Application Number
- CN202510794208.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-13
- Publication Date
- 2025-10-17
AI Technical Summary
Existing surprise recommendation technologies are difficult to achieve accurate modeling, resulting in large recommendation granularity, information cocoons and filter bubbles, and lack of explainability, making them difficult for users to accept.
A multi-task learning framework is used to comprehensively consider multiple surprise factors, calculate the surprise score through a multi-task model, and generate explanatory text based on user comment text to improve the interpretability and accuracy of recommendations.
It achieves fine-grained surprise recommendation optimization, solves the problems of information cocoons and filter bubbles, and improves users' understanding and acceptance of recommendation results.
Smart Images
Figure CN120804434A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data analysis and application, and particularly relates to a surprise degree recommendation method and device, a storage medium and a computer program product. BACKGROUND
[0002] A recommendation system based on data analysis and application shortens search time and relieves information overload by efficiently matching user interests with massive data. However, traditional precise recommendation focuses too much on known interests, causing problems such as filter bubbles and information cocoons. In addition, existing surprise degree recommendation techniques mostly use an optimization paradigm of "relevance + single-dimensional characteristics" (such as novelty, diversity, unexpectedness or usefulness), which can improve the surprise degree of recommendation, but obviously this method has a large recommendation granularity and is difficult to accurately model the surprise degree to achieve accurate surprise degree regulation. Moreover, the consideration is not comprehensive and lacks interpretability, and the recommended results are difficult for users to understand, so they are not easily accepted by users.
[0003] Therefore, there is an urgent need to provide a surprise degree recommendation method, device, storage medium and computer program product to solve at least one of the above problems.
[0004] Application Content
[0005] The present application provides a surprise degree recommendation method, device, storage medium and computer program product.
[0006] In a first aspect, the present application provides a surprise degree recommendation method applied to a product recommendation system, and the method comprises the following steps:
[0007] calculating real values of a plurality of surprise factors according to definitions of the plurality of surprise factors;
[0008] inputting the calculated real values of the plurality of surprise factors into a multi-task model, and obtaining a surprise degree score of a recommended product by using the multi-task model;
[0009] sorting the recommended product according to the surprise degree score of the recommended product, and displaying the recommended product according to a sorting result;
[0010] obtaining a comment text of the recommended product, and performing sentiment preference analysis and keyword extraction on the comment text;
[0011] generating a recommended text explanation by using a pre-trained model according to the extracted keywords; and
[0012] displaying the recommended text explanation to a target user.
[0013] In an embodiment, the plurality of surprise factors include "relevance", "novelty", "divergence", "diversity", "high quality", and "unpopularity", wherein:
[0014] The "relevance" refers to a product similar to the recommended product that the target user has interacted with in the target user's historical behavior;
[0015] The "novelty" refers to a newly listed product in a product database associated with the product recommendation system;
[0016] The "divergence" refers to a product deviating from the target user's historical behavior;
[0017] The "diversity" refers to a product in a category that the target user has not interacted with in the target user's historical behavior;
[0018] The "high quality" refers to a product with a high comprehensive score in the same category as the product currently being searched for;
[0019] The "unpopularity" refers to a product in the product database that is in the long tail distribution.
[0020] In an embodiment, the true value of the plurality of surprise factors is calculated as follows:
[0021]
[0022] wherein Label n represents the true value of the surprise factor n, θ n is a pre-set division threshold for the corresponding surprise factor n,
[0023] The Score n of the "relevance" is calculated as follows:
[0024]
[0025] wherein Score rel (u,p) refers to the Score n , e u is the embedding vector of the target user u, e p is the embedding vector of the recommended product p;
[0026] The true value Score n of the "novelty" is calculated as follows:
[0027]
[0028] wherein Score nov (p) refers to the Score n, current_time is the current timestamp, time(p) is the timestamp of the publication of the recommended product p, and k is a preset adjustable parameter greater than 0;
[0029] The real value Score of the "difference" n is calculated as follows:
[0030]
[0031] where Score diff (u, p) refers to the Score n , e u is the embedding vector of the target user u, and e p is the embedding vector of the recommended product p.
[0032] The real value Score of the "high quality" n is calculated as follows:
[0033] Score qua (p) = average_rating(p),
[0034] where Score qua (u, p) refers to the Score n , and average_rating(p) is the average rating of the recommended product p.
[0035] The real value Score of the "diversity" n is calculated as follows:
[0036]
[0037] where Score div (u, p) refers to the Score n , number of new categories(p) is the number of new categories to which the recommended product p belongs, and total number of categories is the total number of categories to which the recommended product p belongs.
[0038] The real value Score of the "unpopularity" n is calculated as follows:
[0039]
[0040] where Score unpop (u, p) refers to the Score n , and average_rating(p) is the average rating of the recommended product p.
[0041] In an embodiment, the obtaining the serendipity score of the recommended product by using the multi-task model further comprises:
[0042] obtaining a preliminary representation embedding vector X:
[0043] X = e u + e p ,
[0044] wherein e u is an embedding vector of a target user u, and e p is an embedding vector of a target product p;
[0045] inputting the embedding vector X into a task-specific network and a shared network, respectively;
[0046] performing weighted summation on the outputs of the task-specific network and the shared network by using a gating network;
[0047] inputting the result of the weighted summation into a task tower to obtain scores of a plurality of serendipity factors; and
[0048] performing weighted summation on the scores of the plurality of serendipity factors to obtain the serendipity score of the recommended product.
[0049] In an embodiment, the serendipity score Sserendipity is calculated according to the following formula:
[0050] S serendipity = σ1S rel + σ2S nov + σ3S diff + σ4S diy + σ5S qua + σ6S unpop
[0051] wherein σ t (t ∈ [1, 6]) are weight parameters of the serendipity factors “relevance”, “novelty”, “diversity”, “diversity”, “high quality” and “unpopularity”, respectively, S rel , S nov , S diff , S div , S qua , S unpop are scores output by the task tower and corresponding to the serendipity factors “relevance”, “novelty”, “diversity”, “diversity”, “high quality” and “unpopularity”, respectively.
[0052] In an embodiment, the obtaining the review text of the recommended product and performing sentiment preference analysis and keyword extraction on the review text comprises:
[0053] Fine-grained sentiment analysis and preference analysis are performed on the text information of user comments and product comments, and keywords are extracted.
[0054] In an embodiment, the method further includes:
[0055] Analyzing the content corresponding to the plurality of surprise factors and the weight size assigned to obtain an analysis result;
[0056] The generating of the recommended text explanation according to the extracted keywords using the pre-trained model includes:
[0057] The recommended text explanation is generated according to the keywords and the analysis result using the pre-trained model.
[0058] In a second aspect, the present application also provides a surprise degree recommendation system, which includes:
[0059] A memory configured to store computer executable instructions;
[0060] A processor configured to run the computer executable instructions stored in the memory to implement the surprise degree recommendation method described above.
[0061] In a third aspect, the present application also provides a storage medium storing computer executable instructions, which are run by a processing unit to implement the surprise degree recommendation method described above.
[0062] In a fourth aspect, the present application also provides a computer program product including computer executable instructions, which are executed by a processing unit to implement the surprise degree recommendation method described above.
[0063] Compared with the prior art, the surprise degree recommendation method, device, storage medium and computer program product provided by the embodiments of the present application comprehensively consider the relevance, novelty and other factors through a multi-task learning framework, and generate an explanation text in combination with user comment text, aiming to achieve more comprehensive and objective recommendation results, and improve user understanding and acceptance of the recommendations. The multi-dimensional collaborative framework proposed in the present patent realizes fine-grained surprise recommendation optimization while maintaining objective evaluation through six-factor dynamic weight distribution and pre-trained explanation generation, effectively solves the problems of information cocoon, filtering bubbles and other problems in product recommendation, and improves the explainability of the recommendations, thereby improving the understanding and acceptance of the user for the recommendation results. BRIEF DESCRIPTION OF DRAWINGS
[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings.
[0065] Figure 1 is a flowchart of the surprise degree recommendation method provided by an embodiment of the present application;
[0066] Figure 2 is a schematic diagram of an example of the surprise degree recommendation method shown in Figure 1 ;
[0067] Figure 3 is a definition of the concept involved in the method shown in Figure 1 ;
[0068] Figure 4 is a schematic diagram of generating an explanation text provided by an embodiment of the present application.
[0069] Figure 5 is an architecture schematic diagram of the surprise degree recommendation system provided by an embodiment of the present application. DETAILED DESCRIPTION
[0070] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.
[0071] The surprise degree recommendation method of the present application is applied to a surprise degree recommendation device connected with a product database. The surprise degree recommendation method is run to provide more accurate and more easily accepted products by target users according to the products in the product database. The surprise degree recommendation method first uses a multi-task learning model to take multiple surprise factors except randomness as one task respectively, predicts the surprise factor scores of the target users for the candidate products for each task, integrates the obtained multiple scores and generates a surprise degree recommendation result, and then generates an explanation text by using sentiment analysis, pre-training model and other technologies based on the comment text of the users for the recommendation result, so as to improve the explainability and user acceptance of the recommendation.
[0072] Please refer to Figure 1 and Figure 2 , Figure 1is a flowchart of a serendipity recommendation method provided in an embodiment of the present application, Figure 2 is a schematic diagram of an example of the serendipity recommendation method provided in an embodiment of the present application. Figure 1
[0073] In operation S10, real values of the plurality of serendipity factors are calculated according to definitions of the plurality of serendipity factors.
[0074] In operation S11, the real values of the plurality of serendipity factors calculated are input into a multi-task model, and a serendipity score of a recommended product is obtained by using the multi-task model.
[0075] In operation S12, the recommended product is sorted according to the serendipity score of the recommended product, and a recommended display of the recommended product is performed according to a sorting result.
[0076] In operation S13, a keyword is extracted by performing sentiment preference analysis on a comment text of the recommended product.
[0077] In operation S14, a recommended text explanation is generated by using a pre-trained model according to the extracted keyword.
[0078] In operation S15, the recommended text explanation is displayed to a target user.
[0079] In the above operations, some operations can have no sequence, for example, operation S12 and operation S15 can be performed simultaneously. Operation S15 can also be: generating and displaying an icon, and displaying the recommended text explanation according to user operation of the icon, such as single-click or double-click of the icon.
[0080] As shown in Figure 3 The serendipity recommendation method provided in the present application is based on multi-task learning, uses a classic framework PLE framework in multi-task learning as a basic framework of a model, and obtains an embedding representation by analyzing an interactive sequence of a user as an input of a multi-task learning model. In the embodiment, six serendipity components of relevance, novelty, diversity, high quality, and non-popularity of six serendipity factors are regarded as independent tasks, task scores are calculated respectively, and a final serendipity score Sserendipity is obtained by weighted summation.
[0081] As shown in Figure 3 are definitions of six serendipity factors, wherein:
[0082] The serendipity factor "relevance" refers to products similar to a recommended product that are interacted in historical behaviors of a target user. Specifically, for example, the similar products can be defined as products of the same category as the recommended product.
[0083] Surprise factor "novelty" refers to newly listed products in the product database associated with the product recommendation system;
[0084] Surprise factor "diversity" refers to products deviating from the target user's historical behavior, specifically, for example, products not under the same category;
[0085] Surprise factor "diversity" refers to products deviating from the target user's historical behavior, specifically, for example, products not under the same category;
[0086] Surprise factor "diversity" refers to products deviating from the target user's historical behavior, specifically, for example, products not under the same category;
[0087] Surprise factor "non-popularity" refers to products in the long tail distribution in the product database.
[0088] Specifically, in the present embodiment, the true value of the above surprise factor is calculated according to the definition of the above surprise factor, which is the label corresponding to the surprise factor. In the case where there is no labeled data set at present, the label annotation can be completed according to the definition of the surprise factor. According to the preset calculation method, the corresponding score is obtained, and the label is divided in combination with the set hyperparameter threshold. Wherein, Label n represents the true value of surprise factor n, θ n is the preset division threshold corresponding to surprise factor n:
[0089]
[0090] The calculation method of the surprise factor label calculation can be flexibly designed according to the specific definition, and the following provides example references for the calculation of each surprise factor label.
[0091] (1) Relevance: maximize the similarity between the target user u and the recommended product p. Wherein, e u is the embedding vector of the target user u, e p is the embedding vector of the recommended product p.
[0092]
[0093] (2) Novelty: newly appeared products in the product database, which are calculated by using the release timestamp time(p) of the recommended product p. Wherein, current_time is the current timestamp, time(p) is the timestamp of the recommended product p, and k is a preset adjustable parameter used to control the decay rate of novelty score (k>0).
[0094]
[0095] (3) Diversity: products in the product database that deviate from the target user's historical behavior. Among them, e u is the embedding vector of the target user u, e p is the embedding vector of the recommended product p.
[0096]
[0097] (4) Diversity: the recommended product category is diverse, and the category coverage is used to calculate, wherein number of new categories(p) is the number of new categories to which the recommended product p belongs, and total number of categories is the total number of categories to which the recommended product p belongs.
[0098]
[0099] (5) High quality: high-score products in the product database, using the average rating of the recommended product to calculate, wherein average rating(p) is the average rating of the recommended product p.
[0100] Score qua (p)=average_rating(p)
[0101] (6) Unpopularity: products in the long-tail distribution in the product database, calculated using the inverse of popularity. Among them, popularity(p) is the number of interactions of the recommended product p.
[0102]
[0103] Specifically, in the present embodiment, the true value of the plurality of surprise factors calculated is input into a multi-task model, and a recommended surprise score is obtained by using the multi-task model.
[0104] As shown in 2, first, a preliminary representation embedding vector X is obtained by using an existing modeling method:
[0105] X=e u +e p
[0106] Among them, e u is the embedding vector of the target user u, e p is the embedding vector of the target product p. There are many ways to obtain these two embedding vectors, which are described here by taking a matrix decomposition-based collaborative filtering algorithm as an example.
[0107] e u =P T ·x u
[0108] ep = Q T · y p
[0109]
[0110] where P is the user feature matrix, Q is the product feature matrix, both are learnable parameters; x u is the user one-hot encoding vector, y p is the product one-hot encoding vector. The loss function is minimized to reduce the rating prediction error, denotes all observed user-item rating records contained in the dataset, r up is the actual interaction value of the target user u to the target product p, and λ is the regularization coefficient, which is preset by the user to control the regularization strength, are the square of the Frobenius norm of P and Q respectively, to prevent model overfitting.
[0111] The embedding vector X is input into the task-specific network 201 and the shared network 202 respectively:
[0112]
[0113] where, and denote the task-specific network and the shared network respectively, which are generally implemented using a multi-layer perceptron (MLP), K1 is the number of private experts contained in the task-specific network 201, K2 is the number of private experts in the shared network 202, and specifically, K1 = K2 can be set. is the output of the nth task-specific network 201, is the output of the shared network 202, and N represents the number of tasks, which is 6 in this embodiment.
[0114] Then, the gating network 203 is weighted and summed with the outputs of the task-specific network 201 and the shared network 202 to obtain the input Z t of the task tower 204, which is expressed as follows:
[0115]
[0116] where, The weight vector is the output result of the gating network 203 of the nth task, that is, the weight size of the weighting of the nth task expert network and the shared expert network by the nth task, and the value range is 0-1. The value is obtained by calculating the gating network. In this application, the general implementation method is used to complete the function of the gating network 203, that is, the linear layer Linear and the Softmax activation function, and the expression is as follows:
[0117]
[0118] wherein W n is a weight matrix, and b n is a bias vector. n is a weight matrix, and b n is a bias vector.
[0119] Finally, the different task towers 204 output the corresponding different surprise factor scores s n :
[0120] s n =σ(q n (Z n ))
[0121] wherein q n is the task tower of the nth task, implemented by using a multi-layer perception MLP, and sigma is a sigmoid function, which normalizes the result, that is, the adaptation score, to map it to 0-1. A binary cross-entropy loss function is used as the objective function of each task, and the final loss function is defined as:
[0122]
[0123] wherein y is the true value of the nth task, that is, the true value of the nth surprise factor.
[0124] Specifically, in this embodiment, the surprise degree score Sserendipity is calculated according to the following formula:
[0125] S serendipity =σ1S rel +σ2S nov +σ3S diff +σ4S div +σ5S qua +σ6S unpop
[0126] wherein sigma t (t∈[1,6]) are weight parameters of the surprise factors "relevance", "novelty", "diversity", "diversity", "high quality" and "unpopularity", which can be preset according to the importance of the surprise factors in different fields. Srel nov diff div qua unpop are respectively corresponding prediction scores of the surprise factors "relevance", "novelty", "difference", "diversity", "high quality" and "unpopularity" task towers, representing the target user's surprise factor scores of the recommended products. The final surprise degree recommendation result is obtained by sorting the surprise degree scores S serendipity
[0127] In order to improve the generation quality of the explanation text, in the embodiment, the explanation text is generated in combination with the target user's interest preferences and the surprise weight factors in the recommendation process of each field, as shown in Figure 4 First, the text information generated by the user comments, product comments and the like which can directly reflect the user's emotional preferences is subjected to fine-grained sentiment analysis and preference analysis, and the emotional keywords are extracted as one of the important information for the pre-training model 401 to generate the explanation text. Then, the content corresponding to different surprise factors (tasks) and the weight size given are analyzed. In the embodiment, the surprise factors with larger weights are focused on, and the analysis results are used as another important information source for the pre-training model 401 to generate the explanation text. Finally, the above two information sources are integrated by using the existing advanced explanation text generation method to generate high-quality explanation text 402. The generation of the above text can select the existing mode, for example, a template-based method can be used: according to the corresponding explanation template designed for the data set, the key information is embedded into the template to obtain the recommended explanation text. Or use large model fine-tuning, input the key information into the fine-tuned large model to generate reasonable explanation containing key information.
[0128] The application also provides a surprise degree recommendation device, which can be a computer, a server, etc., without limitation. Please refer to Figure 5 As shown, a hardware structure diagram of a surprise degree recommendation device for implementing the surprise degree recommendation method of the present application is shown. The device 50 is connected or communicated with a product database 52. The device 50 includes a processor 501, a memory 503, a communication bus 505, at least one network interface 507 or a user interface 509, which can be a display, a keyboard or a clicking device, for example. The memory 503 can be a high-speed RAM memory or a non-volatile memory such as at least one disk memory. The memory 503 stores execution instructions. When the processor 501 is running, the processor 501 communicates with the memory 503. The processor 501 invokes the instructions stored in the memory 503, i.e., a surprise degree recommendation system 510, to execute the above-mentioned surprise degree recommendation method. The device 50 further includes an operating system 511, which contains various programs for implementing various basic services and processing hardware-based tasks.
[0129] The device 50 provided by the embodiments of the present application can execute the operations included in the above-mentioned surprise degree recommendation method to provide more accurate and more easily accepted products to target users. The implementation principle and technical effects are similar to those of the above-mentioned surprise degree recommendation method, and will not be described here again.
[0130] The embodiments of the present application further provide a computer readable storage medium storing computer executable instructions, which can be the memory 503, for example. The computer executable instructions can enable a processing unit (e.g., the processor 501) to execute the surprise degree recommendation method described in the above-mentioned embodiments. The implementation principle and technical effects are similar to those of the above-mentioned surprise degree recommendation method, and will not be described here again.
[0131] The embodiments of the present application further provide a computer program product including computer executable instructions. When the computer executable instructions are executed by a processing unit (e.g., the processor 501), the surprise degree recommendation method described in the above-mentioned embodiments is implemented. The implementation principle and technical effects are similar to those of the above-mentioned surprise degree recommendation method, and will not be described here again.
[0132] To sum up, compared with the prior art, the surprise degree recommendation method, device, storage medium and computer program product provided by the embodiments of the present application comprehensively consider the correlation, novelty and other factors through a multi-task learning framework, and generate an explanation text in combination with user comment texts, aiming to achieve more comprehensive and objective recommendation results, and improve the understanding and acceptance of the user for the recommendation. The multi-dimensional collaborative framework proposed in the patent realizes fine-grained surprise recommendation optimization while maintaining objective evaluation through six-factor dynamic weight distribution and pre-training explanation generation, effectively solves the problems of information cocoon, filtering bubble and the like in product recommendation, and can improve the explainability of the recommendation, thereby improving the understanding and acceptance of the user for the recommendation results.
[0133] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings of the present application, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.
Claims
1. A surprise recommendation method, applied to a product recommendation system, characterized in that: The method comprises: Calculating the true values of the plurality of surprise factors according to the definitions of the plurality of surprise factors; Inputting the calculated true values of the plurality of surprise factors into a multi-task model, and using the multi-task model to obtain a surprise score for the recommended product; Sorting the recommended products according to their surprise scores, and displaying the recommended products based on the sorting results; Obtaining review texts of the recommended products, and performing sentiment preference analysis and keyword extraction on the review texts; Generate a recommended text explanation using a pre-trained model based on the extracted keywords; and The recommended text is explained and displayed to the target user.
2. The method according to claim 1, characterized in that The plurality of surprise factors include "relevance", "novelty", "difference", "diversity", "high quality" and "non-popularity", wherein: The "relevance" refers to products that the target user has interacted with in his historical behavior and are similar to the recommended product; The “novelty” refers to a product that is newly listed in the product database associated with the product recommendation system; The "difference" refers to products that deviate from the historical behavior of the target users; The "diversity" refers to products that have not been interacted with in the categories that the target user has interacted with in his historical behavior; The term "high quality" refers to products with high comprehensive scores in the same category as the product currently being searched; The “non-popularity” refers to products in the product database that are in the long tail distribution.
3. The method according to claim 2, characterized in that The true values of the multiple surprise factors are calculated as follows: Among them, Label n represents the true value of the surprise factor n, θ n is the pre-set partition threshold corresponding to the surprise factor n, The score of the "relevance" n The calculation is as follows: Score rel (u,p) refers to the Score n , e u is the embedding vector of the target user u, e p is the embedding vector of the recommended product p; The true value of the "novelty" Score n The calculation is as follows: Among them, Score nov (p) refers to the Score n , current_time is the current timestamp, time(p) is the timestamp when the recommended product p is released, and k is a preset adjustable parameter greater than 0; The true value of the "difference" Score n The calculation is as follows: Score diff (u,p) refers to the Score n , e u is the embedding vector of the target user u, e p is the embedding vector of the recommended product p; The true value Score of the "high quality" n The calculation is as follows: Score qua (p)=average_rating(p), Score qua (u,p) refers to the Score n , average_ratint(p) is the average rating of the recommended product p; The true value of the "diversity" Score n The calculation is as follows: Score div (u,p) refers to the Score n , number of categories(p) is the number of new categories to which the recommended product p belongs, and total number of catefories is the total number of categories to which the recommended product p belongs; The true value of "non-popularity" Score n The calculation is as follows: Score unpop (u,p) refers to the Score n , average_rating(p) is the average rating of the recommended product p.
4. The method according to claim 2, characterized in that The obtaining the surprise score of the recommended product by using the multi-task model further includes: Get the initial representation embedding vector X: X=e u +e u , Among them, e u is the embedding vector of target user u, e p is the embedding vector of the target product p; Input the embedding vector X into the task expert network and the shared expert network respectively; Performing a weighted summation of the outputs of the gating network, the task expert network, and the shared expert network; Inputting the weighted summation result into the task tower to obtain the scores of the multiple surprise factors; and The surprise score of the recommended product is obtained by performing a weighted summation on the multiple surprise factor scores.
5. The method according to claim 4, characterized in that The calculation formula for the surprise score Sserendipity is as follows: Among them, σ t (t∈[1,6]) are the weight parameters of the surprise factors “relevance”, “novelty”, “difference”, “diversity”, “high quality” and “non-popularity”, S rel 、S nov 、S diff 、S div 、S qua 、S unpop The scores output by the task tower correspond to the surprise factors "relevance", "novelty", "difference", "diversity", "high quality" and "non-popularity" respectively.
6. The method according to claim 1, characterized in that The obtaining of the review text of the recommended product, and performing sentiment preference analysis and keyword extraction on the review text include: Perform fine-grained sentiment analysis and preference analysis on the text information of user reviews and product reviews and extract the keywords.
7. The method according to claim 6, characterized in that Also includes: Analyze the contents corresponding to the plurality of surprise factors and the weights assigned thereto to obtain analysis results; Generating a recommended text explanation using a pre-trained model according to the extracted keywords includes: The recommended text explanation is generated using the pre-trained model according to the keyword and the analysis result.
8. A surprise recommendation device, characterized in that: The surprise degree recommendation device includes: a memory configured to store computer-executable instructions; A processor, wherein the processor is configured to run the computer executable instructions stored in the memory to implement the surprise degree recommendation method according to any one of claims 1 to 7.
9. A storage medium, characterized in that: The storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processing unit, the surprise degree recommendation method according to any one of claims 1 to 7 is implemented.
10. A computer program product, characterized in that The method comprises computer executable instructions, which, when executed by a processing unit, implements the surprise degree recommendation method according to any one of claims 1 to 7.