An evaluation perception recommendation method considering preference uncertainty

Abstract

The application discloses a review-aware recommendation method considering user preference uncertainty, comprising the following steps: 1. using user reviews and ratings to build a data set, 2. building an asymmetric user preference and product feature representation learning model, and fusing user preference and product features to build a rating prediction model, 3. model training and optimization. The application proposes an asymmetric deep learning recommendation method, which comprehensively considers the difference of user reviews in learning user preference and product representation, and introduces user preference uncertainty, so as to improve the product recommendation performance.

Description

Technical Field

[0001] This invention relates to a recommendation method based on user reviews, particularly for making recommendations when user preferences are uncertain. Background Technology

[0002] With the widespread application of recommender systems across various fields, they not only help consumers reduce search costs but also improve consumer loyalty and spending levels, demonstrating significant commercial value. The most common method is collaborative filtering, but these methods suffer from issues such as data sparsity and cold start problems. As recommender systems evolve, user reviews become a crucial channel for consumers to reflect on their product and service experiences. Therefore, obtaining more comprehensive product attribute knowledge and consumer preferences for attributes through user reviews can improve recommender performance.

[0003] Traditional review-based recommendation methods primarily focus on extracting a series of interpretable variables for prediction, but neglect the rich semantic information inherent in the reviews themselves. Recent research on review-based methods has categorized into document-level and review-level approaches. Document-level recommendation methods struggle to accurately understand user preferences, while review-level methods can obtain more granular semantic features. However, these are expressed according to a symmetrical hierarchical structure and do not consider the different aspects of the product that different users care about. Summary of the Invention

[0004] To address the differences in existing technologies, this invention proposes a review-aware recommendation model that considers uncertain user preferences. This model aims to predict user ratings for products with uncertain preferences, thereby improving product recommendation performance and recommending better products to customers.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The characteristic of this invention, a comment-aware recommendation method considering the uncertainty of user preferences, is that it proceeds according to the following steps:

[0007] Step 1: Collect user reviews and ratings data, as well as product data.

[0008] Constructing a user product rating matrix Where U is the user set, i is the product set, |U| represents the total number of users in user set U, and |I| represents the total number of products in product set i;

[0009] Let user u rate product i as r. ui User u's comment is Among them, C um This represents the m-th comment from user u. u This represents the number of comments posted by user u, where u∈U;

[0010] Let product i be reviewed Among them, C in N represents the nth review for product i. i Let i represent the number of reviews for product i, where i∈I;

[0011] The product index is denoted as id. The index id of each product is vectorized and embedded into a matrix. Where d1 represents the dimension of the index id embedding, and let C be the m-th comment of user u. um The corresponding product index ID embedding vector is denoted as Let the uncertain product preferred by user u be taken as the target product, and let the index ID of the target product be embedded in the vector as follows:

[0012] Step 2: Construct a review perception model that considers the uncertainty of user preferences, including: a user preference learning module, a product representation learning module, and a rating prediction module;

[0013] Step 2.1: Processing of the user preference learning module, including: processing of the text embedding layer, comment representation layer, and comment aggregation layer;

[0014] Step 2.1.1: Processing of the text embedding layer:

[0015] Using the BERT vocabulary embedding matrix W d×|V| With length l um Comment C (mth) um Each word in the vector is encoded into a d-dimensional word embedding vector, which forms an embedding matrix. Where d is the dimension of the vector, and |V| is the total number of words in the BERT vocabulary;

[0016] Step 2.1.2, the comment representation layer for E um After processing, the comment-level features are encoded into comment-level representations O. um ;

[0017] Step 2.1.3, the comment aggregation layer for i um with i target After processing, the preference representation vector X of user u is obtained. u ;

[0018] Step 2.2, processing of the product representation learning module, including: processing of the product feature learning layer and the preference uncertainty aggregation layer;

[0019] Step 2.2.1: The product feature aspect learning layer processes the comments corresponding to product i to obtain the a-th feature aspect level representation Y of product i. i，a ;

[0020] Step 2.2.2: The preference uncertainty aggregation layer represents the preference vector X. u The process is performed to obtain the product representation Y of the user u-target product i pair. i :

[0021] Step 2.3, Prediction Module Processing:

[0022] Step 2.3.1: Use the neural matrix decomposition method to analyze user u's comments C. u And product i's review C i Learning is conducted to obtain user u's preferences p. u Characterization of product i q i ;

[0023] Connect p u X u The final preference representation vector of user u is then obtained.

[0024] Connect q i Y u，i The final representation vector of product i is then obtained.

[0025] Will Composition vector

[0026] Step 2.3.2: Use the neural factor decomposition machine (17) as the prediction function, and use Equation (17) to obtain the prediction result of user u for target product i.

[0027]

[0028] In equation (17), w0 is the global deviation of the input data, w s Let x represent the weight of the s-th feature. s Let represent the s-th feature of χ, and f(·) represent the MLP function used to model higher-order nonlinear feature interactions.

[0029] Step 3: Model training and optimization;

[0030] Step 3.1: Use the scored products for training. Based on the difference between the predicted score and the actual score, calculate the mean square error using equation (18).

[0031]

[0032] In equation (18), L is the generator h. u The number of samples.

[0033] The posterior distribution of the inference network approximation is calculated using equation (19) to compute variational neural inference.

[0034]

[0035] In equation (19), diag represents a diagonal distribution. Indicates average movement. Let represent the standard deviation, and we have:

[0036]

[0037]

[0038] In equations (20) and (21), l3(·) and l4(·) represent linear transformations; k represents the MLP function with activation function, s r Indicates the true rating r ui The embedding vector, where ";" represents the concatenation operator;

[0039] The loss function LOSS for the comment-aware model is constructed using equation (22):

[0040]

[0041] In equation (22), D KL (·) represents the KL divergence, and || represents the distance measurement;

[0042] Step 3.2: Optimize the comment-aware model using the Adam algorithm and calculate the overall loss function LOSS until the loss function LOSS converges or the maximum number of iterations is reached, thereby obtaining the trained comment-aware model.

[0043] Step 4: Use the model to predict the rating of product i for user u who has not rated it and whose preferences are unknown, and make recommendations based on the rating.

[0044] The characteristic of the comment-aware recommendation method considering user preference uncertainty described in this invention is that step 2.1.2 is performed as follows:

[0045] Step 2.1.2.1: Use a convolutional layer on E um The features extracted by the j-th filter at the k-th position are obtained using equation (1).

[0046]

[0047] In equation (1), This represents the j-th filter at the k-th position in the convolutional layer, * represents the convolution operator, and bj This represents the j-th bias term. Let t represent the activation function, t represent the filter width, and k∈(1, l). um -t+1);

[0048] Step 2.1.2.2: Process all features extracted by the j-th filter using a max pooling layer, and then use equation (2) to obtain the m-th comment C. um Comment-level features um，j :

[0049]

[0050] Step 2.1.2.3: Obtain C using J different filters. um The J-th comment-level features are used to classify the m-th comment C using equation (3). um Comment-level features are encoded into comment-level representations O um :

[0051] O um =[o um，1 o um，2 , ..., o um，j , ..., o um，J (3).

[0052] Step 2.1.3 is performed as follows:

[0053] Step 2.1.3.1, put i um with i target As input variables, and using equation (4) to calculate the target product and C um The attention weight of the corresponding product connection

[0054]

[0055] In equation (4), T represents transpose;

[0056] Step 2.1.3.2, using formula (5) After normalization, the normalized attention weight α is obtained. um :

[0057]

[0058] In equation (5), Indicates the calculation of target product i target Comment C from user u (m′th comment) um′ The attention weight of the corresponding inter-product relationships;

[0059] Step 2.1.3.3: Use equation (6) to obtain the preference representation vector X of user u. u :

[0060]

[0061] Step 2.2.1 is performed as follows:

[0062] Step 2.2.1.1: Merge the comments corresponding to product i, and the merged comments should have a length of L. i The words in the comments use a shared embedding vocabulary matrix W d×|V| Transform into a review matrix for product i

[0063] Step 2.2.1.2: Use equation (7) to obtain the projection word matrix related to product i in the a-th feature aspect. Where, d a This represents the dimension of the latent space for the a-th feature.

[0064] D i，a =D i W a (7)

[0065] In equation (7), W a This is represented as the mapping relationship of each word in the review of product i with respect to the a-th feature;

[0066] Step 2.2.1.3: Represent the a-th feature aspect as an embedding vector. Where c is a parameter that controls the text context window;

[0067] Step 2.2.1.4: Use equation (8) to obtain the context vector Z of the e-th word in the review of product i. i,a,e :

[0068]

[0069] In equation (8), (;) represents the concatenation operator; D i,a [e] represents the projected word vector of the e-th word in the review of product i with respect to the a-th feature;

[0070] Step 2.2.1.5: Use equation (9) to measure the attention score α of the e-th word. i,a,e :

[0071]

[0072] Step 2.2.1.6: Use equation (10) to calculate the a-th feature aspect level representation Y of product i. i,a :

[0073]

[0074] Step 2.2.2 is performed as follows:

[0075] Step 2.2.2.1: Calculate user preference X using equation (11). u mean μ u,θ :

[0076] μ u,θ =l1(g θ (X u )) 11

[0077] In equation (11), g θ (·) denotes a multiple perceptron (MLP), l1(·) denotes a linear transformation, and θ denotes the parameters of the Bayesian distribution.

[0078] Step 2.2.2.2: Calculate user preference X using equation (12). u Standard deviation σ u,θ :

[0079]

[0080] In equation (12), l2(·) represents a linear transformation;

[0081] Step 2.2.2.3: Use equation (13) to calculate the interaction h between user u and the various features of the evaluated product. u :

[0082] h u =μ u,θ +|σ u,θ ·∈| (13)

[0083] In equation (13), ∈ represents a random variable that follows a normal distribution;

[0084] Step 2.2.2.4: Use equation (14) to calculate the importance weight of the a-th feature aspect of the target product i to be predicted for user u.

[0085]

[0086] In equation (14), W α W h W y All three parameters are to be trained, where T represents the transpose, tanh is the activation function, and Y is the parameter. i,a This represents the a-th feature aspect of product i at the level of the product i.

[0087] Step 2.2.2.5, Using Equation (15), the importance of the a-th feature aspect of product i for user u normalization α u,i,a :

[0088]

[0089] In equation (15), A represents the number of characteristic aspects;

[0090] Step 2.2.2.6: Use equation (16) to obtain the product representation Y based on the user u-target product i pair. i :

[0091]

[0092] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the evaluation-aware recommendation method, and the processor is configured to execute the program stored in the memory.

[0093] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the evaluation-aware recommendation method.

[0094] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0095] 1. This invention is the first attempt to explore a review-aware recommendation method that takes into account the uncertainty of user preferences. By introducing preference uncertainty into the proposed framework, the degree of uncertainty of user preferences caused by some random factors can be reduced, thereby improving the robustness of the recommendation.

[0096] 2. This invention proposes an asymmetric deep learning recommendation framework. For heterogeneous user reviews, a review-level aggregation strategy is applied to learn user preferences; simultaneously, for homogeneous product reviews, an aspect-based document-level aggregation strategy is used to learn product representations. Integrating Bayesian mechanisms into the framework to model preference uncertainty enables better learning of user preferences and product features, thereby improving product prediction accuracy.

[0097] 3. This invention can provide a reasonable explanation for understanding user preferences and recommending products. It can be used in the task of predicting personalized user needs on e-commerce platforms, and has a wide range of applications and practical value.

[0098] Specific implementation methods

[0099] In this embodiment, a review-aware recommendation method considering user preference uncertainty employs text convolution in the user preference learning module, introducing an attention mechanism to learn user preferences and the user's attention weight for the target product. In the product representation learning module, an ANR method is introduced, using an aspect-level attention mechanism to obtain product representations, and innovatively, user preference uncertainty is introduced to learn the importance of various aspects of the product to the user. Finally, the user preference representation and product representation are integrated, and Neural Factor Decomposition (NFM) is used to predict the potential ratings of products with unknown user preferences, and recommendations are made based on these ratings. Specifically, the method proceeds according to the following steps:

[0100] Step 1: Construct a user product rating matrix Where U is the user set, I is the product set, |U| represents the total number of users in the user set U, and |I| represents the total number of products in the product set I;

[0101] Let user u rate product i as r. ui User u's comment is Among them, C um M represents the m-th comment from user u. u This represents the number of comments posted by user u, where u∈U;

[0102] Let product i be reviewed Among them, C in N represents the nth review for product i. i Let i represent the number of reviews for product i, where i∈I;

[0103] The product index is denoted as id. The index id of each product is vectorized and embedded into a matrix. Where d1 represents the dimension of the index id embedding, and let C be the m-th comment of user u. um The corresponding product index (id) embedding vector is denoted as Let the uncertain product preferred by user u be the target product, and let the index (id) embedding vector of the target product be denoted as .

[0104] Step 2: Construct a review perception model that considers the uncertainty of user preferences, including: a user preference learning module, a product representation learning module, and a rating prediction module;

[0105] Step 2.1: Processing of the user preference learning module, including: processing of the text embedding layer, comment representation layer, and comment aggregation layer;

[0106] Step 2.1.1: Processing the text embedding layer:

[0107] Using the BERT vocabulary embedding matrix W d×|V| With length lum Comment C (mth) um Each word in the vector is encoded into a d-dimensional word embedding vector, which forms an embedding matrix. Where d is the dimension of the vector, and |V| is the total number of words in the BERT vocabulary;

[0108] Step 2.1.2, Processing of the comment representation layer:

[0109] Step 2.1.1.1 Use convolutional layers on E um The features extracted by the j-th filter at the k-th position are obtained using equation (1).

[0110]

[0111] In equation (1), This represents the j-th filter at the k-th position in the convolutional layer, * represents the convolution operator, and b j This represents the j-th bias term. Let t represent the activation function, t represent the filter width, and k∈(1, l). um -t+1);

[0112] Step 2.1.1.2: Process all features extracted by the j-th filter using a max pooling layer, and take the feature with the largest feature value as the feature extracted by the filter. Use equation (2) to obtain the m-th comment C. um Comment-level features um，j :

[0113]

[0114] Step 2.1.1.3: Obtain C using J different filters. um J comment-level features, and thus use equation (3) to classify the m-th comment C um Comment-level features are encoded into comment-level representations O um :

[0115] O um =[o um，1 o um，2 , ..., o um，j , ..., o um，J (3)

[0116] Here, the number of convolutional kernels is set to 40, i.e., J = 40;

[0117] Step 2.1.3, Processing of the comment aggregation layer:

[0118] Step 2.1.3.1: User reviews are heterogeneous; not all information in the reviews can be used to represent user preferences. The relevance of corresponding products is measured using the correlation between product indices. Attention weights are applied to the review-level representations, and a subset of homogeneous reviews is extracted from the heterogeneous review set to learn user preferences related to the target product. (The last part, "i," appears to be a typo and can be left as is.) um with i target As input variables, and using equation (4) to calculate the target product and C um Attention weights for corresponding product-to-product relationships

[0119]

[0120] In equation (4), T represents transpose;

[0121] Step 2.1.3.2, using formula (5) After normalization, the normalized attention weight α is obtained. um :

[0122]

[0123] In equation (5), C represents the m′-th comment of target product i and user u. um′ The attention weight of the corresponding inter-product relationships;

[0124] Step 2.1.3.3: Use equation (6) to obtain the preference representation vector X of user u. u :

[0125]

[0126] Step 2.2, processing of the product representation learning module, including: processing of the product feature learning layer and the preference uncertainty aggregation layer;

[0127] Step 2.2.1, Processing of product features in the learning layer:

[0128] Step 2.2.1.1: Merge the comments corresponding to product i, and the merged comments should have a length of L. i The words in the comments use a shared embedding vocabulary matrix W d×|V| Transform into a review matrix for product i

[0129] Step 2.2.1.2: Use equation (7) to obtain the projection word matrix related to product i in the a-th feature aspect. Where, d a This represents the dimension of the latent space for the a-th feature.

[0130] D i，a =D i W a (7)

[0131] In equation (7), W a This is represented as the mapping relationship of each word in the review of product i with respect to the a-th feature;

[0132] Step 2.2.1.3: Represent the a-th feature aspect as an embedding vector. Where c is a parameter that controls the text context window; in this example, c is set to 3.

[0133] Step 2.2.1.4: Use equation (8) to obtain the context vector Z of the e-th word in the review of product i. i，a，e :

[0134]

[0135] In equation (8), (;) represents the concatenation operator; D i，a [e] represents the projected word vector of the e-th word in the review of product i with respect to the a-th feature.

[0136] Step 2.2.1.5: Use equation (9) to measure the attention score α of the e-th word. i，a，e :

[0137]

[0138] Step 2.2.1.6: Use equation (10) to calculate the a-th feature aspect level representation Y of product i. i，a :

[0139]

[0140] Each aspect-specific subspace has a dimension of 32, and the number of aspects in each dataset is tested from 2 to 10. The latent factor dimension is quantified as [8, 16, 32, 64].

[0141] Step 2.2.2, Processing of the Preference Uncertainty Aggregation Layer:

[0142] Step 2.2.2.1: Calculate user preference X using equation (11). u mean μ u，θ :

[0143] μ u，θ =l1(g θ (X u (11)

[0144] In equation (11), g θ(·) denotes a multiple perceptron (MLP), l1(·) denotes a linear transformation, and θ denotes the parameters of the Bayesian distribution.

[0145] Step 2.2.2.2: Calculate user preference X using equation (12). u Standard deviation σ u，θ :

[0146]

[0147] In equation (12), l2(·) represents a linear transformation;

[0148] Step 2.2.2.3: Use equation (13) to calculate the interaction h between user u and the various features of the evaluated product. u ,

[0149] h u =μ u，θ +|σ u，θ ·∈| (13)

[0150] In equation (13), ∈ represents a random variable that follows a normal distribution; h u Generated by a Bayesian distribution, it describes user preferences as the range or distribution of uncertain values ​​for various aspects of the product. It uses a diagonal Gaussian distribution to construct the solution, applies the reparameterization technique, and performs back random propagation, which can better enhance the robustness of the prediction.

[0151] Step 2.2.2.4: Use equation (14) to calculate the importance weight of the a-th feature aspect of target product i for user u.

[0152]

[0153] In equation (14), W α W h W y All three parameters are to be trained, where T represents the transpose, tanh is the activation function, and Y is the parameter. i，a This represents the a-th feature aspect of product i at the level of the product i.

[0154] Step 2.2.2.5: Use equation (15) to obtain the normalized importance α of the a-th feature aspect of product i for user u. u，i，a :

[0155]

[0156] In equation (15), A is the number of features, and in this example, A is set to 100;

[0157] The preference uncertainty aggregation layer uses equation (16) to obtain the product representation Y based on the user u-target product i pair. u，i :

[0158]

[0159] Step 2.3, Prediction Module Processing:

[0160] Step 2.3.1: User preference X learned from comments u and product characterization Y u，i Each includes rich semantic information representing both the user and the product. However, some user or product-related features are not revealed in the reviews but can be learned from rating patterns. The review C for user u is analyzed using neural matrix factorization. u And product i's review C i To learn and obtain user u's preferences p u Characterization of product i q i ;

[0161] Connect p u X u The final preference representation vector of user u is then obtained.

[0162] connect Product i was subsequently obtained target The final representation vector

[0163] Will Composition vector

[0164] Step 2.3.2: Use the neural factor decomposition machine (17) as the prediction function, and use Equation (17) to obtain the prediction result of user u for target product i.

[0165]

[0166] In equation (17), w0 is the global deviation of the input data, w s Let x represent the weight of the s-th feature. s express The s-th feature, f(*) represents the MLP function used to model higher-order nonlinear feature interactions.

[0167] Step 3: Model training and optimization;

[0168] Step 3.1, Experimental Details:

[0169] This invention selects review data from six categories on Amazon: "Pavilion & Garden," "Office Products," "Digital Music," "Groceries & Food," "Toys & Games," and "Mobile Phones & Accessories," as well as review data from Yelp 2019. Products and users with fewer than five reviews from Yelp 2019 were removed. The data first requires preprocessing: Step 1: Remove distracting words and non-English vocabulary from the reviews; Step 2: Remove meaningless words used in user connections; Step 3: Replace all words with lowercase and use intercept or padding methods to complete the reviews, ensuring that the length and number of reviews cover 85% of users and products.

[0170] The training set (80%), validation set (10%), and test set (10%) were randomly split, and the hyperparameters of the reference method were tuned using a grid search method to obtain the best performance.

[0171] For the PLG, OP, and DM datasets, the batch size is set to 128, while for other datasets it is set to 256.

[0172] The test learning rate is in the range [0.001, 0.002, 0.01], and the test L... 2 The regularization coefficients are in the range [0.0, 0.0001, 0.001, 0.01]. To avoid overfitting, the dropout parameter is set to 0.2 in the hidden layer and 0.5 in the output layer. Training is performed with epoch = 20. The model performance is tested on the validation set every 100 steps.

[0173] Step 3.2: Based on the difference between the predicted score and the actual score, calculate the mean square error using equation (18).

[0174]

[0175] In equation (18), L is the generator h. u The number of samples.

[0176] The posterior distribution of the inference network approximation is calculated using equation (19) to compute variational neural inference.

[0177]

[0178] In equation (19), diag represents a diagonal distribution. Expressing expectations, Let represent the standard deviation, and we have:

[0179]

[0180]

[0181] In equations (20) and (21), l3(·) and l4(·) represent linear transformations; S represents an MLP function with an activation function. r Indicates the true rating r ui The embedding vector, where ";" represents the concatenation operator;

[0182] The loss function LOSS for the comment-aware model is constructed using equation (22):

[0183]

[0184] In equation (22), D KL (*) represents the KL divergence, and || represents the distance measurement;

[0185] Step 3.3: Optimize the comment-aware model using the Adam algorithm and calculate the overall loss function LOSS until the loss function LOSS converges or the maximum number of iterations is reached, thus obtaining the trained comment-aware model.

[0186] Step 4: Use the model to predict the rating of product i for user u who has not rated it and whose preferences are unknown, and make recommendations based on the rating.

[0187] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0188] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0189] In summary, the present invention combines document-level and comment-level methods to design a novel comment-level representation aggregation mechanism for filtering out evaluated products that are irrelevant to the target product. It also incorporates aspect-based document-level methods to extract product feature aspects and designs a new attention method to aggregate product representations. Furthermore, it integrates a Bayesian mechanism into a deep learning neural network to model user preference uncertainty, thereby improving product recommendation performance.

Claims

1. A comment-aware recommendation method considering the uncertainty of user preferences, characterized in that, The procedure is as follows: Step 1: Collect user reviews and ratings data, as well as product data. Constructing a user product rating matrix ,in, For user sets, For product collection, Represents user set The total number of, For product collection The total number in; Allow users For the product The rating is recorded as This allows users The comments are ,in, Indicates user The 1 comment, Indicates user Number of comments published ∈ ; Make products The comments are ,in, Indicates product The 1 comment, Indicated as product Number of comments ∈ ; The product index is denoted as id. The index id of each product is vectorized and embedded into a matrix. ,in, This is represented as the dimension embedded in the index ID, allowing users to... The 1 comment The corresponding product index ID embedding vector is denoted as , with users Let the product with uncertain preference be the target product, and let the index ID of the target product be embedded in the vector. ; Step 2: Construct a review perception model that considers the uncertainty of user preferences, including: a user preference learning module, a product representation learning module, and a rating prediction module; Step 2.1: Processing of the user preference learning module, including: processing of the text embedding layer, comment representation layer, and comment aggregation layer; Step 2.1.1: Processing of the text embedding layer: Using BERT vocabulary embedding matrix The length is The 1 comment Each word in the code is encoded as Word embedding vectors and constructing an embedding matrix ,in, It is the dimension of the vector. This represents the total number of words in the BERT vocabulary. Step 2.1.2, the comment representation layer... After processing, the comment-level features are encoded into comment-level representations. ; Step 2.1.3, the comment aggregation layer... and After processing, the user is obtained. Preference representation vector ; Step 2.2, processing of the product representation learning module, including: processing of the product feature learning layer and the preference uncertainty aggregation layer; Step 2.2.1: The product feature learning layer learns about the product. The corresponding comments are processed to obtain the product. The Characteristic aspect-level representation ; Step 2.2.2: The preference uncertainty aggregation layer represents the preference vector. Process and obtain user information. -Target Products Product characterization : Step 2.3, Prediction Module Processing: Step 2.3.1: Use the neural matrix decomposition method to analyze the user data. Comments and products Comments Learn, and thus gain users preferences With products The representation ; connect After that, the user The final preference representation vector ; connect Based on users -Target Products Product characterization After that, the product was obtained. The final representation vector ; Will , Composition vector ; Step 2.3.2: Use the neural factor decomposition machine of equation (17) as the prediction function, and use equation (17) to obtain the user's... For target products Prediction results : (17) In equation (17), The global deviation of the input data. Indicates the first Each feature weight, express The One characteristic, This represents the MLP function used to model high-order nonlinear feature interactions; Step 3: Model training and optimization; Step 3.1: Use the scored products for training. Based on the difference between the predicted score and the actual score, calculate the mean square error using equation (18). : (18) In equation (18), Is generating users Interactions between various features of the evaluated products Number of samples; The posterior distribution of the inference network approximation is calculated using equation (19) to compute variational neural inference. : (19) In equation (19), Indicates a diagonal distribution. Indicates average movement. Let represent the standard deviation, and we have: (20) (21) In equations (20) and (21), Represents a linear transformation; This refers to an MLP function with an activation function. Indicates the true rating The embedding vector, where ";" represents the concatenation operator; The loss function LOSS for the comment-aware model is constructed using equation (22): (22) In equation (22), It is the KL divergence. Indicates distance measurement; Step 3.2: Optimize the comment-aware model using the Adam algorithm and calculate the overall loss function LOSS until the loss function LOSS converges or the maximum number of iterations is reached, thereby obtaining the trained comment-aware model. Step 4: Use the model to target users Unrated products with unknown preferences Make predictions based on the ratings and then make recommendations based on those ratings.

2. The comment-aware recommendation method considering user preference uncertainty according to claim 1, characterized in that, Step 2.1.2 is performed as follows: Step 2.1.2.1: Use convolutional layers to... After processing, the first equation is obtained using equation (1). The position of the first Features extracted by each filter : (1) In equation (1), Represents the first convolutional layer The position of the first There are several filters, where * denotes the convolution operator. Indicates the first One bias term, This represents the activation function. Indicates the width of the filter. ; Step 2.1.2.2: Use the max pooling layer on the first... All features extracted by the i-th filter are processed, and the i-th feature is obtained using equation (2). 1 comment Comment-level features : (2) Step 2.1.2.3, Use Different filters are obtained of The comment-level features are used to utilize equation (3) to... 1 comment Comment-level features are encoded into comment-level representations : (2)。 3. The comment-aware recommendation method considering user preference uncertainty according to claim 2, characterized in that, Step 2.1.3 is performed as follows: Step 2.1.3.1, will and As input variables, and using equation (4) to calculate the target product and The attention weight of the corresponding product connection : (4) In equation (4), Indicates transpose; Step 2.1.3.2, using formula (5) Normalization is performed to obtain the normalized attention weights. : (5) In equation (5), Indicates the calculation of the target product With users The 1 comment The attention weight of the corresponding inter-product relationships; ; Step 2.1.3.3: Use formula (6) to obtain the user's... Preference representation vector : (6)。 4. The comment-aware recommendation method considering user preference uncertainty according to claim 3, characterized in that, Step 2.2.1 is performed as follows: Step 2.2.1.1: Place the product The corresponding comments will be merged, and the merged length will be [length value missing]. The words in the comments use a shared embedded vocabulary matrix. Convert to product Comment matrix ; Step 2.2.1.2: Obtain the product using formula (7). In the Projection word matrix related to each feature aspect ,in, Indicates the first The dimension of the latent space for each feature; (7) In equation (7), Indicated as product Each word in the comments is in the first Mapping relationships of features; Step 2.2.1.3, the first Each feature aspect is represented as an embedding vector. ,in, These are parameters that control the text context window; Step 2.2.1.4: Obtain the product using formula (8). The first comment Vectors generated from the context of each word : (8) In equation (8), (;) represents the concatenation operator; Indicates product The first comment The word in the first Projected word vectors for each feature aspect; Step 2.2.1.5: Use formula (9) to measure the first... Attention score per word : (9) Step 2.2.1.6: Calculate the product using formula (10). The Characteristic aspect-level representation : (10)。 5. The comment-aware recommendation method considering user preference uncertainty according to claim 4, characterized in that, Step 2.2.2 is performed as follows: Step 2.2.2.1: Calculate user expression preferences using equation (11). mean : (11) In equation (11), This represents a multiple perceptron (MLP). Represents a linear transformation; The parameters representing the Bayesian distribution; Step 2.2.2.2: Use equation (12) to calculate user expression preferences. Standard deviation : (12) In equation (12), Represents a linear transformation; Step 2.2.2.3: Calculate the user using formula (13). Interactions between various features of the evaluated products : (13) In equation (13), ε represents a random variable that follows a normal distribution; Step 2.2.2.4: Use formula (14) to calculate the target product to be predicted. The These characteristics are relevant to users. Importance weight : (14) In equation (14), All three parameters are to be trained. This indicates transpose, and tanh is the activation function. For products The Each feature aspect is represented at the level of representation; Step 2.2.2.5, Use product (15) The These characteristics are relevant to users. The importance of normalization : (15) In equation (15), A represents the number of characteristic aspects; Step 2.2.2.6: Use equation (16) to obtain the user-based... -Target Products Product characterization : (16)。 6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the comment-aware recommendation methods of claims 1-5, and the processor is configured to execute the programs stored in the memory.

7. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the comment-aware recommendation method according to any one of claims 1-5.

Images