Method for predicting product rating based on hierarchical text interaction

Abstract

This invention discloses a product rating prediction method based on hierarchical text interaction, comprising: identifying users and products; obtaining historical review texts and latent feature vectors for users and products; obtaining feature vectors corresponding to the historical review texts; obtaining initial user preference features and initial product attribute features based on the feature vectors; fusing the feature vectors corresponding to the user's historical review texts and the initial product attribute features to obtain user interaction preference features; fusing the feature vectors corresponding to the product's historical review texts and the user's initial preference features to obtain product interaction attribute features; fusing the initial user preference features and the user interaction preference features to obtain user preference features; fusing the initial product attribute features and the product interaction attribute features to obtain product attribute features; and modeling the interaction between user preference features and product attribute features to obtain the predicted user rating for the product. This invention can improve the accuracy of product rating prediction and achieve better recommendation results.

Description

Technical Field

[0001] This invention relates to the field of Internet information classification technology, and in particular to a product rating prediction method based on hierarchical text interaction. Background Technology

[0002] With the continuous development and expansion of e-commerce, the number and variety of goods have also increased rapidly. This means users spend a significant amount of time selecting the products they need. Finding suitable goods in a shorter time has become the future direction of online shopping. Recommendation services based on massive data analysis and data mining technologies have therefore become one of the technologies used by major e-commerce websites.

[0003] Recommender systems are a typical information filtering tool. They capture users' interests and preferences by recording historical interactions between users and products, allowing them to prioritize and present products that users are more likely to be interested in. A recommendation problem can generally be described as a rating prediction problem between users and products; that is, given the rating records already completed between users and products, predicting missing rating records and recommending highly rated products to the user accordingly.

[0004] In practical applications, recommendation systems face a severe data sparsity problem because most users only have historical interaction records with a very small number of products compared to the vast number of products on online platforms. To address this problem, existing rating prediction methods improve accuracy by mining semantic information from review text, thereby enhancing recommendation performance. However, these methods still have limitations in text modeling and text interaction. For text modeling, existing methods simply concatenate all text into a single long text, ignoring the semantic differences between different texts. For text interaction, they postpone mining interaction information until the prediction stage, failing to capture the complex interactions between users and products, resulting in low actual rating prediction accuracy. Summary of the Invention

[0005] To address some or all of the technical problems existing in the prior art, this invention provides a product rating prediction method based on hierarchical text interaction.

[0006] The technical solution of the present invention is as follows:

[0007] A product rating prediction method based on hierarchical text interaction is provided, the method comprising:

[0008] Identify the user and product to be predicted, obtain at least one historical comment text of the user and at least one historical comment text of the product, and obtain the latent feature vectors of the user and the product;

[0009] The historical comment texts are processed using a pre-trained convolutional neural network to obtain the word-level convolutional feature vectors corresponding to each historical comment text;

[0010] The word-level convolutional feature vectors corresponding to the user's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the user's historical comment texts. Similarly, the word-level convolutional feature vectors corresponding to the product's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the product's historical comment texts.

[0011] Based on the feature vectors corresponding to the user's historical review texts and the feature vectors corresponding to the product's historical review texts, obtain the user's initial preference features and the product's initial attribute features;

[0012] The feature vectors corresponding to the user's historical review texts and the initial attribute features of the product are fused to obtain user interaction preference features;

[0013] The feature vectors corresponding to the historical review texts of the product and the user's initial preference features are fused to obtain the product's interaction attribute features;

[0014] The user's initial preference features and the user's interaction preference features are fused to obtain user preference features;

[0015] The product's initial attribute features and the product's interactive attribute features are fused to obtain the product's attribute features;

[0016] The interaction between the user preference features and the product attribute features is modeled to obtain the user's predicted rating for the product.

[0017] In some possible implementations, an embedding retrieval operation is used to obtain the latent feature vector of the user and the latent feature vector of the product.

[0018] In some possible implementations, a pre-trained first attention mechanism model is used to process the word-level convolutional feature vectors corresponding to the user's historical review texts, the user's latent feature vectors, and the product's latent feature vectors to obtain the feature vectors corresponding to the user's historical review texts. Similarly, the word-level convolutional feature vectors corresponding to the product's historical review texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the product's historical review texts.

[0019] In some possible implementations, the feature vector corresponding to the user's historical comment text is obtained using the following formula:

[0020]

[0021] The feature vector corresponding to the historical review text of the product is obtained using the following formula:

[0022]

[0023] Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. This represents the word-level convolutional feature vector corresponding to the user's historical comment text. u represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. i v represents the latent feature vector of the user. j d represents the latent feature vector of the product. i d represents the feature vector corresponding to the user's historical comment text. j This represents the feature vector corresponding to the historical review text of a product.

[0024] In some possible implementations, the user's initial preference features and the product's initial attribute features are obtained based on the feature vectors corresponding to the user's historical review texts and the feature vectors corresponding to the product's historical review texts, including:

[0025] The feature vectors corresponding to the user's historical comment texts are processed using the max pool operation to obtain the user's initial preference features;

[0026] The feature vector corresponding to the historical review text of the product is processed using the maximize pooling operation to obtain the initial attribute features of the product.

[0027] In some possible implementations, a pre-trained second attention mechanism model is used to fuse the feature vectors corresponding to the user's historical comment texts and the initial attribute features of the product to obtain user interaction preference features;

[0028] The product interaction attribute features are obtained by fusing the feature vectors corresponding to the historical review texts of the product and the user's initial preference features using a pre-trained third attention mechanism model.

[0029] In some possible implementations, the user interaction preference features are obtained using the following formula:

[0030]

[0031] The product interaction attribute features are obtained using the following formula:

[0032]

[0033] Among them, v2, W v W vu and b u q represents the parameters of the second attention mechanism model, tanh represents the nonlinear transformation function, and q represents the parameters of the second attention mechanism model. j d represents the initial attribute characteristics of the product. i,k Let represent the feature vector corresponding to the k-th historical comment text of a user, m represent the total number of historical comment texts of the user, and s represent the feature vector. i Representing user interaction preference characteristics, v1, W u W uv and b v p represents the parameters of the third attention mechanism model. i d represents the user's initial preference characteristics. j,t Let represent the feature vector corresponding to the t-th historical review text of the product, where n represents the total number of historical review texts of the product, and t j This represents the interactive attributes and characteristics of a product.

[0034] In some possible implementations, a pre-trained gating network mechanism model is used to fuse the user's initial preference features and the user's interaction preference features to obtain user preference features, and the product's initial attribute features and the product's interaction attribute features are fused to obtain product attribute features.

[0035] In some possible implementations, a pre-trained feedforward neural network is used to model the interaction between the user preference features and the product attribute features to obtain the user's predicted rating for the product.

[0036] In some possible implementations, the user's predicted rating for the product is obtained using the following formula:

[0037]

[0038] in, This represents the user's predicted rating for the product, where L represents the number of layers in the feedforward neural network, and φ... l Let h represent the l-th layer of the neural network. l-1 σ represents the output of the (l-1)th layer of the neural network. l Let W represent the activation function of the l-th layer of the neural network. l And vector b l Denotes the parameters of the l-th layer of the neural network, [;] denotes the vector concatenation operation, ⊙ denotes the vector dot product operation, D i D represents user preference characteristics. jRepresenting product attribute features, the l-th layer of the neural network outputs h from the (l-1)-th layer. l-1 As input, output h l This is used as the input to the (l+1)th layer of the neural network, and so on, with the last layer outputting a scalar. As a predictive rating of products by users.

[0039] The main advantages of the technical solution of this invention are as follows:

[0040] The product rating prediction method based on hierarchical text interaction of the present invention models the interaction between users and products at the word level and the text level, which can fully explore text interaction information at different granularities and levels, significantly improve the accuracy of product rating prediction, and achieve better recommendation results. Attached Figure Description

[0041] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this invention, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0042] Figure 1 This is a flowchart of a product rating prediction method based on hierarchical text according to an embodiment of the present invention;

[0043] Figure 2 This is a flowchart illustrating the information processing of a product rating prediction method based on hierarchical text, according to an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0045] The technical solutions provided by the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] refer to Figure 1-2 An embodiment of the present invention provides a product rating prediction method based on hierarchical text interaction, the method comprising the following steps S1-S9:

[0047] Step S1: Determine the user and product to be predicted, obtain at least one historical comment text of the user and at least one historical comment text of the product, and obtain the latent feature vectors of the user and product.

[0048] In one embodiment of the present invention, the user's historical review text represents the existing user reviews of all products, including the product to be predicted, and the product's historical review text represents the existing user reviews of that product, including the user to be predicted.

[0049] In one embodiment of the present invention, the number of historical comment texts obtained is determined according to the actual situation. Generally speaking, the more historical comment texts used in the prediction process, the higher the prediction accuracy, but the higher the processing cost and processing time.

[0050] In one embodiment of the present invention, for the user and the product to be predicted, the latent feature vector of the user and the latent feature vector of the product are obtained respectively through the embedding lookup operation.

[0051] Setting: A set of users and products to be predicted i V j > indicates that U i V represents the user. j Indicates the product. User U i And Product V j Through embedding lookup operations, the latent feature vectors of users and products are obtained respectively. i The latent feature vector is embedded in an embedding vector u i It indicates that product V j The latent feature vector is embedded in an embedding vector v j This means that embedding retrieval is a one-to-one mapping, where each user corresponds to a unique embedding vector, and each product corresponds to a unique embedding vector. Embedding vectors can be extracted from various explicit information sources. When there are no explicit information sources, they are learned from the user's interaction records with the product. Therefore, the embedding vector is a model parameter, which is determined through automatic learning and updating during model training.

[0052] The specific operation of embedding retrieval is as follows: Assume that the user embedding matrix can be represented as U∈R M×d Where M is the total number of users and d is the dimension of the embedding vector, then the user embedding matrix U contains Uembed. i The corresponding row vector is U. i Embedding vector u i Similarly, suppose the product embedding matrix can be represented as V∈R N×d Where N is the total number of items and d is the dimension of the embedding vector, then the item embedding matrix V contains V... j The corresponding row vector is V. j Embedded vector v j .

[0053] ​Step S2: Process the historical comment texts using a pre-trained convolutional neural network to obtain the word-level convolutional feature vectors corresponding to each historical comment text.

[0054] In one embodiment of the present invention, a pre-trained convolutional neural network is used to process historical comment texts to obtain word-level convolutional feature vectors corresponding to each historical comment text, further comprising:

[0055] Identify the constituent words of the historical commentary text;

[0056] The word-level convolutional features of each word are obtained using a pre-trained convolutional neural network;

[0057] Based on the word-level convolutional features of each word, determine the word-level convolutional feature vector corresponding to the historical comment text.

[0058] Specifically, the user's historical comment text is represented as D. ui ={D i,1 ,…, i,m The product's historical review text is represented as D. vj ={D j,1 ,…, j,n}, D i,m D represents the m-th historical comment text of a user. j,n This represents the nth historical review text for a product.

[0059] Furthermore, taking a user's historical comment text as an example, let's assume: a historical comment text consists of p words, represented as... w l This represents the l-th word.

[0060] By embedding retrieval operations, p words are mapped one by one to p vectors. Then, a pre-trained convolutional neural network is used to obtain the w of each word. l Phrase-level convolutional features.

[0061] Specifically, the word w is obtained through a convolutional neural network. l The phrase-level convolutional features are represented as follows:

[0062]

[0063] Among them, c 1,f The word w l The word-level convolutional features, where f represents the nonlinear transformation function, W f Denotes the convolution kernel, b f This represents the bias coefficient. The convolution operation selects the word w. l A total of s words are convolved on the left and right sides to form corresponding convolutional features c. l,f .

[0064] Furthermore, assuming the convolutional neural network uses k convolutional kernels for convolution operations, then for each word, a k-dimensional convolutional feature vector is formed, specifically represented as:

[0065]

[0066] Among them, c l The word w l k-dimensional convolutional feature vectors The word w represents the word obtained by the k-th convolution kernel. l Phrase-level convolutional features.

[0067] In one embodiment of the present invention, for each historical comment text of a user and each historical comment text of a product, the above processing can be used to obtain the word-level convolutional feature vector corresponding to each historical comment text, which can effectively capture the word-level information of the comment text.

[0068] The parameters of the convolutional neural network are determined through automatic learning and updating during model training.

[0069] Step S3: Process the word-level convolutional feature vectors, latent feature vectors of users and products corresponding to the user's historical review texts to obtain the feature vectors corresponding to the user's historical review texts; and process the word-level convolutional feature vectors, latent feature vectors of users and products corresponding to the product's historical review texts to obtain the feature vectors corresponding to the product's historical review texts.

[0070] In one embodiment of the present invention, a pre-trained first attention mechanism model is used to process the word-level convolutional feature vectors corresponding to the user's historical review texts, the user's latent feature vectors, and the product's latent feature vectors to obtain the feature vectors corresponding to the user's historical review texts; the pre-trained first attention mechanism model is also used to process the word-level convolutional feature vectors corresponding to the product's historical review texts, the user's latent feature vectors, and the product's latent feature vectors to obtain the feature vectors corresponding to the product's historical review texts.

[0071] Specifically, the settings are as follows: This represents the word-level convolutional feature vector corresponding to the user's historical comment text. u represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. i v represents the latent feature vector of the user. j This represents the latent feature vector of a product.

[0072] The first attention mechanism model processes the word-level convolutional feature vectors corresponding to the user's historical comment text, the user's latent feature vector, and the product's latent feature vector, and represents them as follows:

[0073]

[0074] The first attention mechanism model is used to process the word-level convolutional feature vectors corresponding to the historical review texts of products, the latent feature vectors of users, and the latent feature vectors of products, and represents them as follows:

[0075]

[0076] Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. d i d represents the feature vector corresponding to the user's historical comment text. j This represents the feature vector corresponding to the historical review text of the product. The parameters of the first attention mechanism model are determined through automatic learning and updating during model training.

[0077] In one embodiment of the present invention, for each historical comment text of a user and each historical comment text of a product, the feature vector corresponding to each historical comment text of the user and the feature vector corresponding to each historical comment text of the product are obtained through the above processing.

[0078] Step S4: Based on the feature vectors corresponding to the user's historical review text and the product's historical review text, obtain the user's initial preference features and the product's initial attribute features.

[0079] In one embodiment of the present invention, user initial preference features and product initial attribute features are obtained based on the feature vectors corresponding to the user's historical review text and the feature vectors corresponding to the product's historical review text, including:

[0080] Maximize pooling is used to process the feature vectors corresponding to users' historical comment texts to obtain users' initial preference features;

[0081] Maximize pooling is used to process the feature vectors corresponding to the historical review texts of products to obtain the initial attribute features of the products.

[0082] Specifically, taking m historical user reviews and n historical product reviews as an example, the following formula is used to obtain the initial user preference features and the initial product attribute features:

[0083]

[0084] Where, p iThis represents the initial user preference features, max-pooling represents the maximum pooling operation, and d i,m Let q represent the feature vector corresponding to the m-th historical comment text of a user. j d represents the initial attribute characteristics of the product. j,n This represents the feature vector corresponding to the nth historical review text of a product.

[0085] Step S5: Fuse the feature vectors corresponding to the user's historical review texts and the initial attribute features of the product to obtain user interaction preference features.

[0086] In one embodiment of the present invention, a pre-trained second attention mechanism model is used to fuse the feature vectors corresponding to the user's historical comment texts and the initial attribute features of the product to obtain user interaction preference features.

[0087] Specifically, the feature vectors corresponding to users' historical review texts and the initial attribute features of products are fused using the second attention mechanism model as follows:

[0088]

[0089] Among them, v2, W v W vu and b u q represents the parameters of the second attention mechanism model, tanh represents the nonlinear transformation function, and q represents the parameters of the second attention mechanism model. j d represents the initial attribute characteristics of the product. i,k Let s represent the feature vector corresponding to the user's k-th historical comment text. i This represents user interaction preference features. The parameters of the second attention mechanism model are determined through automatic learning and updating during model training.

[0090] Step S6: Fuse the feature vectors corresponding to the historical review texts of the product with the user's initial preference features to obtain the product's interaction attribute features.

[0091] In one embodiment of the present invention, a pre-trained third attention mechanism model is used to fuse the feature vectors corresponding to the historical review texts of a product with the user's initial preference features to obtain the product's interaction attribute features.

[0092] Specifically, the feature vectors corresponding to the historical review texts of the product and the user's initial preference features are fused using the third attention mechanism model as follows:

[0093]

[0094] Among them, v1 and W u W uv and b vThe parameters of the third attention mechanism model are represented by tanh, which represents the nonlinear transformation function, and p. i d represents the user's initial preference characteristics. j,t Let t represent the feature vector corresponding to the t-th historical review text of a product. j This represents the interactive attribute features of the product. The parameters of the third attention mechanism model are determined through automatic learning and updating during model training.

[0095] Step S7: Fuse the initial user preference features and the user interaction preference features to obtain user preference features.

[0096] In one embodiment of the present invention, a pre-trained gating network mechanism model is used to fuse user initial preference features and user interaction preference features to obtain user preference features.

[0097] Specifically, the gating network mechanism model is used to fuse user initial preference features and user interaction preference features as follows:

[0098]

[0099] Where σ represents the sigmoid nonlinear function, σ(x) = 1 / (1 + exp(-x)), W g1 W g2 and b g Denotes the parameters of the gated network mechanism model, ⊙ represents the vector dot product operation, and D i This represents user preference features. The physical meaning of the gating network mechanism is to automatically determine the weights of two different input vectors, and then fuse the two input vectors through a weighted sum operation. The parameters of the gating network mechanism model are automatically learned and updated during model training.

[0100] Step S8: Fuse the initial attribute features and interactive attribute features of the product to obtain the product attribute features.

[0101] In one embodiment of the present invention, a pre-trained gating network mechanism model is used to fuse the initial attribute features and interactive attribute features of a product to obtain the product attribute features.

[0102] Specifically, the initial attribute features and interactive attribute features of the product are fused and represented using a gating network mechanism model as follows:

[0103]

[0104] Where σ represents the sigmoid nonlinear function, σ(x) = 1 / (1 + exp(-x)), W g1 W g2 and b gDenotes the parameters of the gated network mechanism model, ⊙ represents the vector dot product operation, and D j This indicates the product's attribute characteristics.

[0105] Step S9: Model the interaction between user preference features and product attribute features to obtain the user's predicted rating for the product.

[0106] In one embodiment of the present invention, a pre-trained feedforward neural network is used to model the interaction between user preference features and product attribute features to obtain the user's predicted rating for the product.

[0107] Specifically, the interaction between user preference features and product attribute features is modeled using a feedforward neural network as follows:

[0108]

[0109] in, This represents the user's predicted rating for the product, where L represents the number of layers in the feedforward neural network, and φ... l Let h represent the l-th layer of the neural network. l-1 σ represents the output of the (l-1)th layer of the neural network. l Let W represent the activation function of the l-th layer of the neural network. l And vector b l Let h represent the parameters of the l-th layer of the neural network, [;] represent vector concatenation, and ⊙ represent vector dot product. The l-th layer of the neural network is represented by the output h of the (l-1)-th layer. l-1 As input, output h l This is used as the input to the (l+1)th layer of the neural network, and so on, with the last layer outputting a scalar. This serves as a predicted rating for products from users. The parameters of the feedforward neural network are determined through automatic learning and updating during model training.

[0110] Furthermore, in one embodiment of the present invention, the embedding vectors, parameters of the convolutional neural network, parameters of each attention mechanism model, parameters of the gating network mechanism model, and parameters of the feedforward neural network are all updated and optimized through model training.

[0111] Specifically, in one embodiment of the present invention, the embedding vector, convolutional neural network, various attention mechanism models, gating network mechanism models, and feedforward neural network are trained through the following steps:

[0112] Obtain the training dataset, which includes users and products, historical review texts of users and products, and actual ratings of products by users;

[0113] Using users, products, and their corresponding historical review texts as input, and the actual ratings of users for products as output, the embedding vectors, convolutional neural networks, various attention mechanism models, gating network mechanism models, and feedforward neural networks are trained using the backpropagation algorithm based on a preset objective function.

[0114] Specifically, during model training, the embedding vectors, convolutional neural networks, various attention mechanism models, gating network mechanism models, and feedforward neural networks all have initialization parameters. Based on these initialization parameters and the aforementioned specific data processing procedures and formulas, a corresponding predicted score can be obtained for each input training data. An objective function is calculated based on the actual score and the predicted score. Based on the objective function, the backpropagation algorithm is used to continuously update the parameters of each model until the set iteration stopping condition is reached, thus completing model training and determining the embedding vectors, convolutional neural networks, various attention mechanism models, gating network mechanism models, and feedforward neural networks. The models mentioned include embedding vectors, convolutional neural networks, various attention mechanism models, gating network mechanism models, and feedforward neural networks.

[0115] In one embodiment of the present invention, the objective function is set as follows during model training:

[0116]

[0117] Where L represents the objective function, D represents the training data set, and r ij Indicates the actual score. This indicates the predicted score.

[0118] The objective function defined above represents the error between the predicted score and the actual score. Optimizing this objective function is equivalent to minimizing this error, which is equivalent to improving the model's prediction accuracy.

[0119] Furthermore, in one embodiment of the present invention, the parameters of the model are updated using the following formula:

[0120]

[0121] Where, Θ t+1 Let Θ represent the set of parameters for all models at the (t+1)th iteration. t Let represent the set of parameters for all models at iteration t, Δ[·] represent the optimizer, η represent the learning rate, and Θ represent the set of model parameters. The optimizer can be, for example, Adam or SGD, and the learning rate needs to be preset to control the speed of parameter updates.

[0122] The product rating prediction method based on hierarchical text interaction provided in one embodiment of the present invention models the interaction between users and products at the word level and the text level, which can fully explore text interaction information at different granularities and levels, significantly improve the accuracy of product rating prediction, and achieve better recommendation results.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Additionally, the terms "front," "back," "left," "right," "upper," and "lower" in this document refer to the placement shown in the accompanying drawings.

[0124] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A product rating prediction method based on hierarchical text interaction, characterized in that, include: Identify the user and product to be predicted, obtain at least one historical comment text of the user and at least one historical comment text of the product, and obtain the latent feature vectors of the user and the product; The historical comment texts are processed using a pre-trained convolutional neural network to obtain the word-level convolutional feature vectors corresponding to each historical comment text; The word-level convolutional feature vectors corresponding to the user's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the user's historical comment texts. Similarly, the word-level convolutional feature vectors corresponding to the product's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the product's historical comment texts. Based on the feature vectors corresponding to the user's historical review texts and the feature vectors corresponding to the product's historical review texts, obtain the user's initial preference features and the product's initial attribute features; The feature vectors corresponding to the user's historical review texts and the initial attribute features of the product are fused to obtain user interaction preference features; The feature vectors corresponding to the historical review texts of the product and the user's initial preference features are fused to obtain the product's interaction attribute features; The user's initial preference features and the user's interaction preference features are fused to obtain user preference features; The product's initial attribute features and the product's interactive attribute features are fused to obtain the product's attribute features; Model the interaction between the user preference features and the product attribute features to obtain the user's predicted rating for the product; The historical comment text is processed using a pre-trained convolutional neural network to obtain the word-level convolutional feature vector corresponding to each historical comment text, including: determining the constituent words of the historical comment text; obtaining the word-level convolutional features of each word using a pre-trained convolutional neural network; and determining the word-level convolutional feature vector corresponding to the historical comment text based on the word-level convolutional features of each word. The first attention mechanism model is pre-trained to process the word-level convolutional feature vectors corresponding to the user's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors to obtain the feature vectors corresponding to the user's historical comment texts. Similarly, the word-level convolutional feature vectors corresponding to the product's historical comment texts, the user's latent feature vectors, and the product's latent feature vectors are processed to obtain the feature vectors corresponding to the product's historical comment texts. Based on the feature vectors corresponding to the user's historical review text and the product's historical review text, the user's initial preference features and the product's initial attribute features are obtained, including: The feature vectors corresponding to the user's historical comment texts are processed using the max pool operation to obtain the user's initial preference features; The feature vector corresponding to the historical review text of the product is processed using the max pool operation to obtain the initial attribute features of the product; The user's interaction preference features are obtained by fusing the feature vectors corresponding to the user's historical review texts and the initial attribute features of the product using a pre-trained second attention mechanism model. The feature vectors corresponding to the historical review texts of the product and the user's initial preference features are fused using a pre-trained third attention mechanism model to obtain the product's interaction attribute features; The user's initial preference features and the user's interaction preference features are fused using a pre-trained gating network mechanism model to obtain user preference features; and the product's initial attribute features and the product's interaction attribute features are fused to obtain product attribute features. A pre-trained feedforward neural network is used to model the interaction between the user preference features and the product attribute features to obtain the user's predicted rating for the product.

2. The product rating prediction method based on hierarchical text interaction according to claim 1, characterized in that, The latent feature vectors of the user and the product are obtained by using an embedding retrieval operation.

3. The product rating prediction method based on hierarchical text interaction according to claim 1, characterized in that, The feature vector corresponding to the user's historical comment text is obtained using the following formula: The feature vector corresponding to the historical review text of the product is obtained using the following formula: Where W and b represent the parameters of the first attention mechanism model, and tanh represents the nonlinear transformation function. This represents the word-level convolutional feature vector corresponding to the user's historical comment text. u represents the phrase-level convolutional feature vector corresponding to the historical review text of a product. i v represents the latent feature vector of the user. j d represents the latent feature vector of the product. i d represents the feature vector corresponding to the user's historical comment text. j This represents the feature vector corresponding to the historical review text of a product.

4. The product rating prediction method based on hierarchical text interaction according to claim 3, characterized in that, The user interaction preference features are obtained using the following formula: The product interaction attribute features are obtained using the following formula: Among them, v2, W v W vu and b u q represents the parameters of the second attention mechanism model, tanh represents the nonlinear transformation function, and q represents the parameters of the second attention mechanism model. j d represents the initial attribute characteristics of the product. i,k Let represent the feature vector corresponding to the k-th historical comment text of a user, m represent the total number of historical comment texts of the user, and s represent the feature vector. i Representing user interaction preference characteristics, v1, W u W uv and b v p represents the parameters of the third attention mechanism model. i d represents the user's initial preference characteristics. j,t Let represent the feature vector corresponding to the t-th historical review text of the product, where n represents the total number of historical review texts of the product, and t j This represents the interactive attributes and characteristics of a product.

5. The product rating prediction method based on hierarchical text interaction according to claim 4, characterized in that, The user's predicted rating for the product is obtained using the following formula: in, This represents the user's predicted rating for the product, where L represents the number of layers in the feedforward neural network, and φ... l Let h represent the l-th layer of the neural network. l-1 σ represents the output of the (l-1)th layer of the neural network. l Let W represent the activation function of the l-th layer of the neural network. l And vector b l Denotes the parameters of the l-th layer of the neural network, [;] denotes the vector concatenation operation, ⊙ denotes the vector dot product operation, D i D represents user preference characteristics. j Representing product attribute features, the l-th layer of the neural network outputs h from the (l-1)-th layer. l-1 As input, output h l This is used as the input to the (l+1)th layer of the neural network, and so on, with the last layer outputting a scalar. As a predictive rating of products by users.

Images