A product matching method and device, electronic equipment and storage medium

Abstract

The application discloses a product matching method and device, electronic equipment and a storage medium, and relates to the technical field of product matching. The method comprises the following steps: obtaining historical comment information of a user and product introduction information of a product to be matched; determining user granularity features in the historical comment information and product granularity features in the product introduction information according to a preset granularity; and fusing the user granularity features and the product granularity features to determine the matching degree of the product to be matched and the user. The application can improve the accuracy of user feature extraction, enhance the precision of user demand and interest preference, expand the product range, improve the matching degree of the product and the user, and improve the user experience.

Description

Technical Field

[0001] This invention relates to the field of computer application technology, and in particular to a product matching method, apparatus, electronic device, and storage medium. Background Technology

[0002] The rapid development and widespread application of big data technology have propelled my country's digital transformation, but the massive amounts of data generated have inevitably led to a serious "information overload" problem. To help users select items that match their interests from this vast amount of data, it's necessary to match products that meet their needs with them, thereby reducing the difficulty for users to obtain products. Currently, a common method is to extract user needs and interests based on user profiles and historical behavior, then combine this with product information to determine the user's level of interest in the product, thus achieving product matching. However, in practice, because users have typically interacted with only a small fraction of products, the matching accuracy for other products is often inaccurate, and the matching results are often unsatisfactory when faced with entirely new products. Furthermore, every user review is treated equally during the product matching process, resulting in indistinct extracted text features and preventing users from finding matching products, significantly impacting the user experience. Summary of the Invention

[0003] This invention provides a product matching method, apparatus, electronic device, and storage medium to improve the accuracy of user feature extraction, enhance the precision of user needs and interests, expand the product range, improve the matching degree between products and users, and enhance the user experience.

[0004] According to one aspect of the present invention, a product matching method is provided, wherein the method includes: acquiring a user's historical review information and product description information of a product to be matched; determining user granularity features in the historical review information and product granularity features in the product description information according to a preset granularity; and fusing the user granularity features and the product granularity features to determine the degree of matching between the product to be matched and the user.

[0005] According to another aspect of the present invention, a product matching device is provided, wherein the device comprises:

[0006] The information acquisition module is used to acquire users' historical review information and product introduction information of the products to be matched; the feature extraction module is used to determine user granular features in the historical review information and product granular features in the product introduction information according to a preset granularity; the product matching module is used to fuse the user granular features and the product granular features to determine the degree of matching between the product to be matched and the user.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0008] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the product matching method according to any embodiment of the present invention.

[0009] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the product matching method according to any embodiment of the present invention.

[0010] The technical solution of this invention collects users' historical review information and product introduction messages, determines the user granularity features of historical review information and the product granularity features of product introduction information according to different preset granularities, and determines the matching degree between the product to be matched and the user by fusing the user granular features and product granular features. This improves the accuracy of user feature extraction, enhances the precision of user needs and interests, expands the product range, and improves the matching degree between the product and the user, thereby improving the user experience.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a product matching method provided according to Embodiment 1 of the present invention;

[0014] Figure 2 This is a flowchart of another product matching method provided according to Embodiment 2 of the present invention;

[0015] Figure 3 This is an example diagram of a product matching method provided according to Embodiment 3 of the present invention;

[0016] Figure 4 This is an example diagram of a comment-level feature extraction method provided in Embodiment 3 of the present invention;

[0017] Figure 5a This is an example diagram illustrating the similarity between a product and a user according to Embodiment 3 of the present invention;

[0018] Figure 5b This is an example diagram illustrating the similarity between a product and a user according to Embodiment 3 of the present invention;

[0019] Figure 6 This is a schematic diagram of the structure of a product matching device according to Embodiment 4 of the present invention;

[0020] Figure 7 This is a schematic diagram of the structure of an electronic device that implements the product matching method of this invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Example 1

[0024] Figure 1 This is a flowchart of a product matching method according to Embodiment 1 of the present invention. This embodiment is applicable to situations where a user and a product are matched. This method can be executed by a product matching device, which can be implemented in hardware and / or software, and can be configured in a terminal device or server. Figure 1 As shown, the method includes:

[0025] Step 110: Obtain the user's historical review information and the product introduction information of the product to be matched.

[0026] Historical reviews can be user ratings of different products over a period of time. These reviews can include textual reviews and ratings, and the time period for which they are attributed can be determined through experience or experimentation. The products to be matched can be products used to match users. These products can be physical or virtual. Product descriptions can include reviews and descriptions of the products to be matched, and these descriptions can come from manufacturers or other users.

[0027] In this embodiment of the invention, historical review information of users over a period of time and product introduction information of one or more products to be matched can be collected. It is understood that the range of products to be matched can be automatically set by the system or selected according to product characteristics.

[0028] Step 120: Determine user granularity features in historical review information and product granularity features in product introduction information according to preset granularity.

[0029] The preset granularity can be the hierarchical granularity of feature extraction, and can include aspects, comments, etc. In some embodiments, the preset granularity can be set manually by the user. User granular features can be user features extracted based on historical comment information at different preset granularities, while product granular features can be user features extracted from product description information at different preset granularities.

[0030] In this embodiment of the invention, features of historical comment information and product introduction information can be extracted at different preset granularities as user-granular features and product-granular features. It is understood that user-granular features and product-granular features can include feature information extracted at at least two preset granularities. Here, the method of extracting user-granular features and product-granular features is restricted. For example, bag-of-words model, N-Grams bag model, TF-IDF model, document similarity, etc. can all be used to extract features at different preset granularities.

[0031] Step 130: Integrate the granular features of each user and the granular features of the product to be matched to determine the degree of matching between the product and the user.

[0032] The fusion process can involve combining user-level features at different preset granularities with product-level features at different preset granularities. In some embodiments, it can combine user-level features at the comment level with user-level features at the aspect level. The fusion can be achieved using methods such as neural network models, expert systems, and attention mechanism models.

[0033] In this embodiment of the invention, user-level features and product-level features can be fused separately, and then the matching degree between the user and the product to be matched can be determined based on the fused user-level features and product-level features. In some embodiments, the geometric distance between the feature vectors of the fused user-level features and product-level features can be used as the similarity with the product to be matched; the higher the similarity, the higher the matching degree between the product and the user. In other embodiments, an aspect-based neural recommender (ANR) model can be used to determine the matching degree between the product and the user.

[0034] In this embodiment of the invention, by collecting users' historical review information and product introduction messages, the user granularity features of the historical review information and the product granularity features of the product introduction information are determined according to different preset granularities. By fusing the user granularity features and the product granularity features, the matching degree between the product to be matched and the user is determined, thereby improving the accuracy of user feature extraction, enhancing the precision of user needs and interests, expanding the product range, and improving the matching degree between the product and the user, which can improve the user experience.

[0035] Example 2

[0036] Figure 2 This is a flowchart of another product matching method provided by Embodiment 2 of the present invention. This embodiment is a specific modification based on the above-described embodiments. See also... Figure 2 The method provided in this embodiment of the invention specifically includes the following steps:

[0037] Step 210: Read historical comment information from the first preset database according to the user's identification information.

[0038] The first preset database can be a database that stores users' historical comment information, and the first preset database can provide an information access interface.

[0039] In this embodiment of the invention, user identification information can be retrieved. This identification information can be the user's unique identification information, which may include user ID, user phone number, etc. Historical comment information can be retrieved in a first preset database according to the user's identification information through the information access interface.

[0040] Step 220: Find the product identifier that matches the user's identifier information.

[0041] Among them, the product identifier can be the unique identification information of the product to be matched.

[0042] In this invention, product identifiers can correspond to user identifiers, and this correspondence can be preset. For example, users with different permissions can have different products to be matched. Product identifiers with corresponding relationships can be found based on the extracted user identifiers.

[0043] Step 230: Read the product introduction information from the second preset database according to the product identifier.

[0044] The second preset database may be a database that stores product introduction information. In some embodiments, the first preset database and the second preset database may be the same database.

[0045] In this embodiment of the invention, the information access interface of the second preset database can be invoked to search for product introduction information in the second preset database according to the product identifier.

[0046] Step 240: Divide the historical review information into at least one historical review information aspect set according to the product aspect to which it belongs, and divide the product introduction information into at least one product introduction information set according to the product aspect to which it belongs.

[0047] The historical review information aspect set can be a dataset divided according to aspects. This set can be used to extract user-granular features for each aspect. In some embodiments, the historical review information aspect set may only include historical review information for each aspect; in other embodiments, it may consist of user-granular features based on the review hierarchy. Similarly, the product introduction information set can be used to extract user-granular features based on the aspect hierarchy. Product aspects can be aspects extracted from user characteristics through analysis, and may include price, performance, service, product type, etc.

[0048] In this embodiment of the invention, historical review information can be divided into one or more historical review information aspect sets and product introduction information can be divided into one or more product introduction information sets for different product aspects. The product aspects corresponding to different historical review information aspect sets may be different, and the product aspects corresponding to different product introduction information sets may also be different.

[0049] Step 250: Extract first-granularity user features corresponding to the comment granularity from historical comment information and extract first-granularity product features corresponding to the comment granularity from product introduction information.

[0050] In this embodiment of the invention, features can be extracted from historical comment information and product introduction information at the comment granularity, and the extracted features can be recorded as first-granularity user features and second-granularity user features, respectively.

[0051] Step 260: Extract second-granularity user features at the corresponding aspect granularity from the historical review information set and extract second-granularity product features at the corresponding aspect granularity from the product introduction information set.

[0052] In this embodiment of the invention, features can be extracted from the historical review information aspect set and the product introduction information set according to aspect granularity. The extracted feature information can be denoted as the second-granularity user feature and the second-granularity product feature, respectively.

[0053] Step 270: Use the first-granularity user feature and the second-granularity user feature as user-granularity features, and use the first-granularity product feature and the second-granularity product feature as product-granularity features.

[0054] Step 280: Correct the user granularity features and product granularity features according to the preset attention mechanism text set.

[0055] The preset attention mechanism text set can include keywords and their weights. The preset attention mechanism can adjust the importance of user-level features and product-level features, making them more consistent with the actual situation.

[0056] In this embodiment of the invention, a preset attention mechanism text set can be searched, and the importance of user-level features and product-level features can be adjusted according to the weights of keywords within the preset attention mechanism text set. In some embodiments, the weight coefficients of different keywords within the preset attention mechanism text set can be extracted, and the weight coefficients can be used to adjust user-level features and product-level features.

[0057] Step 290: Determine the similarity between the second user granularity feature within the user granularity feature and the second product granularity feature within the product granularity feature.

[0058] The similarity can be defined as the degree of similarity between the second user-level features and the second product-level features at the aspect level. The similarity can be determined based on the distance between the similarity matrices of the second user-level features and the second product-level features. This distance can be cosine distance, Euclidean distance, Manhattan distance, BM25 similarity, Jaccard distance, etc.

[0059] In this embodiment of the invention, feature matrices corresponding to the second user-level features and the second product-level features can be generated respectively. The similarity can be determined by the distance between the feature matrices. In some embodiments, the distance between the feature matrices can be determined by cosine distance, Euclidean distance, Manhattan distance, BM25 similarity, Jaccard distance, etc.

[0060] Step 2100: Generate user importance features corresponding to user granular features and product importance features corresponding to product granular features based on similarity.

[0061] User importance features can better reflect user characteristics. These features can be determined based on similarity; for example, features with higher similarity are considered more important. In some embodiments, the similarity value can be adjusted to reflect the value of the user-level granular feature, making the user's historical granular feature a user importance feature. Product importance features can be generated based on product-level features that reflect importance, and these features are generated based on similarity.

[0062] In this embodiment of the invention, user-level features and product-level features can be adjusted according to similarity. The adjusted features can be used as user importance features and product importance features. In some embodiments, the adjustment method may include finding the similarity value corresponding to each feature for user-level features and product-level features, and using the product of the weight coefficient corresponding to the similarity value and the user-level features and product-level features as user importance features and product importance features.

[0063] Step 2110: Read the preset latent semantic model matching formula, and determine the degree of matching based on user granularity features, product granularity features, user importance features, and product importance features according to the preset latent semantic model matching formula.

[0064] Among them, the preset latent semantic model matching formula can be a matching formula generated based on the latent semantic model (LFM).

[0065] In this embodiment of the invention, a preset latent semantic model matching formula can be read, and user granularity features, product granularity features, user importance features, and product importance features can be substituted into the preset latent semantic model matching formula respectively. The degree of matching can be determined by the calculation of the formula.

[0066] In this embodiment of the invention, historical review information is searched according to user identifiers to determine product identifiers matching the user identifiers. Product description information corresponding to the product identifiers is then searched. Historical review information and product description information are respectively divided into a historical review information aspect set and a product description information set according to product aspects. First-granularity user features and first-granularity product features are extracted from the historical review information and product description information at the review granularity level. Second-granularity user features and second-granularity product features are extracted from the historical review information aspect set and product description information set at the aspect granularity level. A preset attention mechanism text set is used to correct the user-granularity features and product-granularity features to determine the... The similarity between second-level user features and second-level product features is used to convert these features into user importance features and product importance features, respectively. Then, according to a pre-defined latent semantic model matching formula, the matching degree between the user and the product to be matched is determined using these user-level features, product-level features, user importance features, and product importance features. By fusing user-level features and product-level features to determine the matching degree between the product and the user, the accuracy of user feature extraction is improved, the precision of user needs and interests is enhanced, the product range is expanded, and the matching degree between the product and the user is increased, thus improving the user experience.

[0067] Furthermore, based on the above embodiments of the invention, obtaining the user's historical review information and the product introduction information of the product to be matched also includes:

[0068] The rating value of the user corresponding to the product to be matched is determined according to the preset neural network model and the historical review information;

[0069] Find auxiliary users who have the same rating as the user for the product to be matched;

[0070] The auxiliary comment information of the auxiliary user is filled into the historical comment information.

[0071] In this embodiment of the invention, a preset neural network model can be invoked to determine the historical review information and the user's rating value for the product to be matched. Based on the rating value, auxiliary users with the same or similar rating values ​​for the product to be matched can be found. The auxiliary user's auxiliary review information can also be added to the historical review information to change the sparsity of the user granular feature extraction.

[0072] Furthermore, based on the above embodiments of the invention, the preset latent semantic model matching formula includes:

[0073]

[0074] Among them, a u，m Identify the user granularity feature, a i，mIdentifying the particle size characteristics of the product, β u，m Identify the user importance feature, β i，m Identify the important characteristics of the product, b u b i b o The user's bias, the product to be matched, and the global bias are identified respectively.

[0075] Example 3

[0076] Figure 3 This is an example diagram of a product matching method according to Embodiment 3 of the present invention. In this embodiment, an MCRA model is provided to determine the matching between a user and a product to be matched. Two parallel neural networks are constructed within this model to model the user's review text and the item's review text, respectively, jointly learning the latent features of the user and the item. One neural network uses the user's written reviews to build a user interest preference model, while the other neural network builds an item feature model based on the reviews received by the item. Since the two neural networks perform essentially the same processing at the same layer, this patent mainly describes the processing details of the user review text neural network. See also... Figure 3The product matching method provided in this embodiment of the invention specifically includes: (1) an embedding layer. The task is to convert text information into low-dimensional text semantic vectors. This layer processes user comment text and item ID information respectively to obtain the word vector matrix of user comment text and the vector representation of item ID. (2) a comment-level processing layer. The task is to learn the importance weight of each comment text to the user. This layer includes two parts: a comment-level encoding layer and a comment-level attention mechanism. First, the CNN convolution operation is used in the comment-level encoding layer to obtain the representation of each user comment text. Then, the attention mechanism is introduced in combination with the item ID vector to learn the weight of each comment text in a remote supervision manner. (3) an aspect-level processing layer. The task is to learn the representation of each user comment text under different aspect categories. The CNN convolution operation is used in the aspect-level encoding layer to obtain the potential feature vector of each word combined with the context. The aspect extraction is performed using the gating mechanism to obtain the correlation between word features and each aspect. Then, the aspect is further extracted using the global attention mechanism to obtain the aspect representation of each user's own comment text. Finally, the aspect representation after auxiliary comment processing is fused to obtain the final aspect representation of each user's comment. (4) a multi-level fusion layer. The task is to fuse comment-level semantic information with aspect-level semantic information. The aspect representation of each comment text is weighted and summed using the weights of the user comment text to obtain the user's aspect representation. (5) Aspect Importance Evaluation Layer. The task is to obtain the dynamic aspect importance of the user. For the target user-item pair, a collaborative attention mechanism is introduced to analyze the aspect correlation of the user-item pair to obtain the aspect importance of the user-item pair. (6) Rating Prediction Layer. The task is to predict the rating of the target user-item pair. The rating is predicted using the aspect representations of the user and item, as well as the aspect importance, using a biased LFM algorithm.

[0077] Furthermore, the embedding layer uses Word2vec word vector technology and One-hot encoding to process the user comment text and item ID information respectively, to obtain the word vector matrix of the user comment text and the embedding vector of the item ID.

[0078] Step 1: Vectorize the comment text. Process each comment from user u. Given the text of the i-th comment from user u... w j For comments The j-th word, l is the comment length, padded with 0s if insufficient. An embedding matrix pre-trained on Word2vec is used in the embedding layer. Where V represents the vocabulary size and d represents the word embedding dimension, the words in the comment text are mapped to word vectors to obtain the comment text. The word vector matrix D u，i = [w1, w2, ..., w j , ..., w l], Let be the word vector of the j-th word.

[0079] Step 2: ID vectorization, the text of user u's i-th comment. The one-hot encoded vector i′ of the corresponding item ID u，i After being input into the embedding layer, the embedding vector i can be obtained. u，i As shown in equation (3-1).

[0080] i u，i =W id i′ u，i (3-1)

[0081] in, It is a one-hot encoded vector. Let d' be the item embedding weight matrix, d′ be the dimension of the item ID embedding vector, and |I| be the total number of items. Similarly, we can obtain the embedding representation u of the user ID corresponding to the j-th comment of item i. i，j .

[0082] Furthermore, in the comment-level processing layer, after obtaining the word embedding matrix of the comments, the importance of each user / item comment text to itself is obtained using the comment-level attention mechanism. The feature extraction process within the comment-level processing layer is as follows: Figure 4 As shown.

[0083] Step 1: Comment Level Encoding Layer. The text of the i-th comment from user u is processed using the TextCNN method. The word vector matrix D of the comment text is obtained in the embedding layer. u，i Then, n convolutional filters with a sliding window size of s are used to learn the comment text representation that incorporates local context features, as shown in Equation (3-2):

[0084]

[0085] * indicates a convolution operation. Let D be the convolution weight matrix of the j-th convolutional filter. u，i [h:h+s-1] is matrix D u，i Slices within the sliding window starting at position h, This is a bias term. It is filter f j Local contextual features extracted from a sliding window starting at position h.

[0086] After obtaining local context features using filters, max pooling (Equation (3-3)) is used to retain the most important context features, and all retained features are concatenated (Equation (3-4)) as the output of the max pooling layer. Finally, a fully connected layer is passed to obtain the final representation of the comment text as shown in Equation (3-5):

[0087]

[0088] o = [o1, o2, ..., o n (3-4)

[0089]

[0090] in, It is a vector of n max-pooled values ​​concatenated. It is the text vector of user u's i-th comment processed by a Convolutional Neural Network (CNN). It is a weight matrix. It is the bias vector.

[0091] Step 2: Comment-level attention mechanism. An attention mechanism is introduced to help remotely supervise the learning of the importance of each comment from user u. A two-layer network is used to calculate the attention weight p′ for user u's i-th comment. u，i Input user u and the feature vector o of the i-th comment u，i The embedded representation of the item being commented on, i u，i The weight calculation process for the attention mechanism is shown in equations (3-6) and (3-7):

[0092] p′ u，i =h T ReLU(W o o u，i +W i i u，i +b1)+b2 (3-6)

[0093] in, b2 is the model parameter, and t1 is the size of the hidden layer of the attention network.

[0094] The attention scores are normalized using the softmax function to obtain the final weight of the comment text, i.e., the contribution of the i-th comment to user u's preference, as shown in formula (3-7):

[0095]

[0096] Where k is the number of comments by user u. Thus, the weight of each comment by user u is obtained, denoted as p.u，1 p u，2 , ..., p u，k For the item neural network, performing the same operation as above yields the weight q for each comment on item i. u，1 q u，2 ,...,q u，k′ , where k′ is the number of comments on item i.

[0097] Furthermore, the aspect-level processing layer mainly extracts aspect category information contained in the comment text to better represent user interests and item characteristics.

[0098] Step 1: Obtain the word vector matrix D of user u's i-th comment in the embedding layer. u，i Next, the latent feature vector of the j-th word combined with its context is extracted, and the semantic features extracted from the j-th word are concatenated using n convolutional filters with a window size of s. First, it is necessary to supplement the beginning and end of the comment word vector matrix. The zero vectors, for the k-th filter, are derived from matrix D. u，i Extract the context features of the j-th word. Specific details are shown in formula (3-8):

[0099]

[0100] Where * represents a convolution operation. Let be the convolution weight matrix of the k-th convolutional filter. It is a sliding window pair matrix D with the center word located at j and the window size being s. u，i Matrix slices, This is the bias term. n filters with different convolutional weights extract different contextual features for each word and its local context (i.e., s consecutive words), resulting in the latent feature vector of the j-th word combined with its context, as shown in formula (3-9):

[0101]

[0102] Therefore, the feature matrix C of the i-th comment of user u is synthesized. u，i =[c u，i，1 c u，i，2 c u，i，l ]. It is the latent feature vector of the j-th word.

[0103] Step 2: Aspect feature extraction, latent features c of the j-th word in user u's i-th comment. u，i，jThis can be viewed as a combination of multiple aspects. Further, a gating mechanism is used to determine the correlation between features and each aspect. Referring to the gating control method introduced by GLU, the activation function σ in the second term on the right-hand side of the formula is used as a soft switch to control the correlation between latent features and aspects. Therefore, for the m-th aspect, the feature vector extracted from the j-th word for the m-th aspect can be expressed as formula (3-10).

[0104]

[0105] Where σ is the sigmoid activation function, and ⊙ represents the element-wise multiplication operation at the corresponding position. Let m represent the transformation matrix of the m-th aspect. Let t2 represent the bias vector of the m-th aspect. t2 is the latent dimension of the aspect representation. Through the above process, we can obtain the contextual features G of user u's i-th comment regarding specific words across the M aspects. u，i These features were used for further aspect extraction:

[0106] G u，i，m =[g u，i，1，m g u，i，2，m , ..., g u，i，l，m (3-11)

[0107] G u，i =[G u，i，1 G u，i，2 , ..., G u，i，M (3-12)

[0108] Step 3: Aspect-level global attention mechanism. Reviews in different domains typically focus on different aspects. For example, the book domain tends to include plot and characters, while the film domain tends to include actors and special effects. Therefore, the model generates an aspect-level global attention matrix by truncating a normal distribution: V = [v1, v2, ..., v...]. M ], During training, the matrix values ​​are gradually corrected. The aspect global matrix V is used to guide aspect extraction. Specifically, a u，i，m From G u，i，m Extract the representation of the m-th aspect, as shown in formula (3-13):

[0109]

[0110]

[0111] β u，i，j，m Let j represent the importance of word j to the m-th aspect. Therefore, we obtain the representation of the M aspects of user u's i-th comment, forming A. u，i =[a u，i，1 a u，i，2, ..., a u，i，M ],

[0112] Step 4, auxiliary comments, utilizes auxiliary comments from similar users to alleviate the sparsity of user-generated comment data. A similar user, for a given user-item pair, is another user who has given the same rating to the target user for that item. The process of finding similar user auxiliary comments includes searching for auxiliary comments and selecting auxiliary comments. Searching for auxiliary comments includes:

[0113] Input: User u, Item i u，i

[0114]

[0115] rate = GetRate(u, i u，i #Get user u's view on item i u，i rating

[0116] Select auxiliary comments

[0117] if

[0118]

[0119] end if

[0120] if

[0121]

[0122] end if

[0123] return

[0124] Selected supplementary comments include:

[0125] Enter item i u，i Rating r

[0126] set = GetUsers(i u，i ,r)#Get item i u，i user set for r

[0127] Review = GetReview(i u，i ,set)#Get the comment set corresponding to the user set

[0128] review = Max(qI) u，i ,Review)#Get the comment with the highest weight in the item review

[0129] Return to review

[0130] Because users behave differently from similar users, the features of two reviews are also heterogeneous. The CARL model argues that stacking a CNN on the context matrix is ​​effective for rating prediction and performance consistency, especially when the text semantics are inconsistent. Therefore, for review text... Using the aspect-level coding layer method described earlier, as shown in equations (3-8) and (3-9), i.e., combining the contextual convolution operation, we obtain... Based on this, an abstraction layer of CNN based on max pooling is added to extract higher-level semantic features. The specific operation is shown in equations (3-15)-(3-17):

[0131]

[0132]

[0133]

[0134] Where * represents a convolution operation. The convolution weight matrix is... This is a bias term. j This is the max-pooling output of the j-th filter. It is a sliding window pair matrix with the center word located at h and the window size being s. The matrix slices are then used to obtain the word vector matrix for the auxiliary comments. The same operation is performed after obtaining the word vector matrix from the user's own comments, i.e., formula (3-10)-(3-14) to obtain the aspect representation of the auxiliary comments.

[0135] Since different user-item pairs are related to different user interests, in order to extract effective features from auxiliary comments while filtering out irrelevant features, the user comment aspect representation and the auxiliary comment aspect representation are merged together using the gating mechanism of inter-element interaction to obtain the final aspect representation of user u's i-th comment, as shown in formula (3-19).

[0136]

[0137]

[0138] in, The circle (⊙) represents the concatenation operation, and the circle (⊙) represents the element-wise multiplication operation. It is a transformation matrix. It is the bias vector.

[0139] Furthermore, through the multi-level fusion layer and the aspect-level processing layer, the aspect features A of user u's i-th comment can be obtained. u，i Considering the different semantic granularities at different levels and the varying effectiveness of each comment, the model utilizes comment-level attention mechanisms to obtain comment weights p. u，i The aspect features of user u are obtained by weighted summation of the aspect representations of k comments, as shown in formula (3-20).

[0140]

[0141] Therefore, by integrating the aspect features of all user u's comments, we obtain the aspect features A of user u considering the semantic information of the comment level. u =[a u，1 a u，2 , ..., a u，M Similarly, we can obtain aspect features A of item i. i .

[0142] Furthermore, in the aspect importance evaluation layer, referencing the dynamic interaction mechanism in the ANR model, a collaborative attention mechanism is introduced to learn the dynamic aspect importance of users and items. When learning user aspect importance, the aspect-level features of items are used as context, and vice versa. To calculate user aspect importance while considering item aspect-level features, it is first necessary to know how the target user-item pair matches at the aspect level. Here, user u aspect features A are used. u and the characteristics of item i A i Using formula (3-21), we obtain an aspect-level similarity matrix S:

[0143]

[0144] It is a learning weight matrix and a similarity matrix. Each item in the table represents the similarity between the target user and the item in the corresponding aspect.

[0145] Next, the similarity matrix S is used as a feature, and Equations (3-22) and (3-23) are used to evaluate the importance of the user-item pair at the aspect level.

[0146]

[0147]

[0148] t3 is the learning parameter, and t3 is the size of the hidden layer for evaluating importance. These represent the aspect importance of user u and item i in the M aspects of evaluation, respectively.

[0149] Furthermore, the rating prediction layer combines aspects of users and items to represent A. u =[a u，1 a u，2 , ..., a u，M A i =[a i，1 a i，2 , ..., a i，M Importance of aspects β u =[β u，1 ,β u，2 , ..., β u，M ]、β i =[β i，1 ,β i，2 , ..., β i，M Using a biased LFM model, the target item-object pair rating can be predicted, and the calculation formula is (3-24):

[0150]

[0151] b u b i b o These are the user, the item, and the global bias.

[0152] In this embodiment of the invention, in addition to the model used in the method provided by this embodiment, other benchmark models can also be used, including the rating collaborative filtering model FM, the classic deep learning model based on comment DeepCoNN for document modeling, the DeepCoNN model PARL that integrates auxiliary comments, the NARRE model that considers the importance of each comment, the HUITA model that considers the multi-granular semantics of comment text, the ALFM model based on aspect-level topic modeling, and the ANR model based on aspect-level deep learning modeling, etc. The product matching effects of different benchmark models are shown in the table below:

[0153] Table 1. Comparison of model performance on four different domain datasets

[0154]

[0155] In the table, the best results are indicated in bold for each dataset, and the second-best results are indicated by underscores. Analysis of the experimental results in the table leads to the following conclusions: the MCRA model performs optimally on all four datasets. Compared to the second-best results, the model's MSE performance metrics are improved by 0.47%, 1.1%, 2.0%, and 0.78% on the four datasets, respectively. The MCRA model consistently outperforms the FM, DeepCoNN, PARL, NARRE, HUITA, ALFM, and ANR models, demonstrating the effectiveness of multi-granularity collaborative attention recommendation models incorporating auxiliary comments in predicting ratings for recommender systems.

[0156] The product matching method provided in this embodiment of the invention generates a matching program that is interpretable. It randomly selects a user-item pair (User 1-Item 1 and User 2-Item 2) from the Movies&TV and Toys&Games datasets, respectively, as the case analysis objects. Based on the aspect similarity matrix S of the user-item pair, it obtains the attention scores between various aspects of the user-item pair, such as... Figure 5a and Figure 5b The presentation shows the attention scores for each aspect of the user and each aspect of the item in two user-item pairs. Figure 5a This shows the correlation between user 1 and item 1, while Figure 5b The matrix shows the correlation between user 2 and item 2. The pair of aspects with the highest scores in the matrix represents the most closely related user-item pair. It can be seen that aspect 1 of user 1 is most closely related to aspect 1 of item 1, and aspect 4 of user 2 is most closely related to aspect 3 of item 2.

[0157] Next, we find the top five words with the highest attention scores in the close aspect category among user comments, item comments, and auxiliary comments, and display sentences containing these words (the parts of the words in the comments are bolded). The attention score of a word is obtained by combining comment-level attention and aspect-level attention. For the j-th word in the i-th comment of user u, the attention score of word j is shown in formula (3-25).

[0158] score j =p u，i β u，i，j，m (3-25)

[0159] Where p u，i β is the attention weight of user u's i-th comment. u，i，j，mLet $j$ be the importance of word $j$ in user $u$'s i-th comment to the $m-th aspect. The table below shows the top 5 words with the highest attention scores extracted from user comments, item comments, and auxiliary comments for each user-item pair in a specified aspect category, sorted from highest to lowest attention score. The comments containing the words with the highest attention scores are then organized, and labels are manually assigned to their respective aspects based on word attributes.

[0160] Table 2: List of keywords related to user-item attention to specific categories

[0161]

[0162] Observing the cases in Table 3, the first example is a user 1-item 1 pair extracted from the Movies&TV dataset. User 1 and item 1 have the highest correlation in terms of "genre". Analyzing the sentences containing words with high attention scores reveals that, based on user reviews and auxiliary reviews, user 1 prefers "laughing" and "humorous" movies, while item 1, a movie about a rock band, is described as "extremely funny," which matches user 1's viewing habits. Therefore, the model predicts a high rating, and the accuracy of the predicted rating is verified by actual reviews and real ratings.

[0163] The second example involves analyzing a user-item 2 pair extracted from the Toys & Games dataset. Through word sets and reviews, it can be inferred that user 2 focuses on the "quality" of the item, particularly its durability ("durable"). The user reviews also indicate dissatisfaction with items lacking durability. Item 2 is described as a toy that is "not as strong," leading to the conclusion that user 2 will be dissatisfied with item 2. The model predicts a low rating of 3.97. Observation of actual reviews and real ratings validates the accuracy of the predicted rating.

[0164] Table 3. Comments containing words with high user-item attention scores.

[0165]

[0166]

[0167] Example 4

[0168] Figure 6 This is a schematic diagram of a product matching device according to Embodiment 4 of the present invention. Figure 6 As shown, the device includes: an information acquisition module 301, a feature extraction module 302, and a product matching module 303.

[0169] The information acquisition module 301 is used to acquire users' historical comment information and product introduction information of the products to be matched.

[0170] The feature extraction module 302 is used to determine user-level features in the historical comment information and product-level features in the product introduction information according to a preset granularity.

[0171] Product matching module 303 is used to fuse the user granularity features and the product granularity features to determine the degree of matching between the product to be matched and the user.

[0172] In this embodiment of the invention, the information acquisition module collects users' historical review information and product introduction messages. The feature extraction module determines the user-level features of the historical review information and the product-level features of the product introduction information according to different preset granularities. The product matching module determines the matching degree between the product to be matched and the user by fusing the user-level features and the product-level features. This improves the accuracy of user feature extraction, enhances the precision of user needs and interests, expands the product range, and improves the matching degree between products and users, thereby improving the user experience.

[0173] Furthermore, based on the above embodiments of the invention, the information acquisition module 301 includes:

[0174] The comment search unit is used to read the historical comment information from a first preset database according to the user's identification information.

[0175] The product matching unit is used to find product identifiers that match the user's identifier information.

[0176] The product search unit is used to read the product description information from a second preset database according to the product identifier.

[0177] Furthermore, based on the above embodiments, the information acquisition module 301 further includes:

[0178] The auxiliary search unit is used to determine the rating value of the user for the product to be matched according to a preset neural network model and the historical review information; find auxiliary users who have the same rating value as the user for the product to be matched; and fill the historical review information with the auxiliary user's auxiliary review information.

[0179] Furthermore, based on the above embodiments of the invention, the feature extraction module 302 includes:

[0180] A classification unit is used to divide the historical review information into at least one historical review information aspect set according to the product aspect to which it belongs, and to divide the product introduction information into at least one product introduction information set according to the product aspect to which it belongs.

[0181] The comment feature unit is used to extract first-granularity user features corresponding to the comment granularity for the historical comment information and to extract first-granularity product features corresponding to the comment granularity for the product introduction information.

[0182] An aspect feature unit is used to extract a second granular user feature corresponding to the aspect granularity for the historical review information aspect set and to extract a second granular product feature corresponding to the aspect granularity for the product introduction information set; the first granular user feature and the second granular user feature are used as the user granular feature, and the first granular product feature and the second granular product feature are used as the product granular feature.

[0183] Furthermore, based on the above embodiments, a feature correction module is also included, which is used to correct the user-granular features and the product-granular features according to a preset attention mechanism text set.

[0184] Furthermore, based on the above embodiments of the invention, the product matching module 303 includes:

[0185] A similarity unit is used to determine the similarity between the second user granularity feature within the user granularity feature and the second product granularity feature within the product granularity feature.

[0186] An importance unit is used to generate user importance features corresponding to the user granularity features and product importance features corresponding to the product granularity features according to the similarity.

[0187] The formula matching unit is used to read a preset latent semantic model matching formula and determine the degree of matching based on the user granularity feature, the product granularity feature, the user importance feature, and the product importance feature according to the preset latent semantic model matching formula.

[0188] Furthermore, based on the above embodiments of the invention, the preset latent semantic model matching formula includes:

[0189]

[0190] Among them, a u，m Identify the user granularity feature, a i，m Identifying the particle size characteristics of the product, β u，m Identify the user importance feature, β i，m Identify the important characteristics of the product, b u b i b o The user's bias, the product to be matched, and the global bias are identified respectively.

[0191] The product matching device provided in the embodiments of the present invention can execute the product matching method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of executing the method.

[0192] Example 5

[0193] Figure 7 This is a schematic diagram of the structure of an electronic device implementing the product matching method of an embodiment of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0194] like Figure 7 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0195] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0196] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as product matching methods.

[0197] In some embodiments, the product matching method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the product matching method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the product matching method by any other suitable means (e.g., by means of firmware).

[0198] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0199] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0200] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0201] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0202] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0203] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0204] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0205] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A product matching method, characterized in that, The method includes: Obtain users' historical review information and product descriptions of the products to be matched; According to a preset granularity, user granularity features are determined in the historical review information and product granularity features are determined in the product introduction information respectively; The user granularity features and the product granularity features are combined to determine the degree of matching between the product to be matched and the user; The preset granularity includes at least aspect granularity and comment granularity. Correspondingly, determining user granularity features in the historical comment information and product granularity features in the product introduction information according to the preset granularity includes: The historical review information is divided into at least one historical review information aspect set according to the product aspect to which it belongs, and the product introduction information is divided into at least one product introduction information set according to the product aspect to which it belongs; Extract first-granularity user features corresponding to the granularity of the comments from the historical comment information, and extract first-granularity product features corresponding to the granularity of the comments from the product introduction information; Extract second-granularity user features corresponding to the granularity of the historical comment information aspect set, and extract second-granularity product features corresponding to the granularity of the product introduction information set; The first granularity user feature and the second granularity user feature are used as the user granularity feature, and the first granularity product feature and the second granularity product feature are used as the product granularity feature. The product aspect refers to the aspects of analyzing and extracting user characteristics, including price, performance, service, and product type.

2. The method according to claim 1, characterized in that, The process of obtaining the user's historical review information and the product description information of the product to be matched includes: The historical comment information is read from a first preset database according to the user's identification information; Find the product identifier that matches the user's identification information; The product information is retrieved from the second preset database according to the product identifier.

3. The method according to claim 1, characterized in that, The process of obtaining the user's historical review information and the product description information of the product to be matched also includes: The rating value of the user corresponding to the product to be matched is determined according to the preset neural network model and the historical review information; Find auxiliary users who have the same rating as the user for the product to be matched; The auxiliary comment information of the auxiliary user is filled into the historical comment information.

4. The method according to claim 1, characterized in that, Also includes: The user-level features and product-level features are modified according to a preset attention mechanism text set.

5. The method according to claim 1 or 4, characterized in that, The process of fusing the user-level granular features and the product-level granular features to determine the matching degree between the product to be matched and the user includes: Determine the similarity between the second user granularity feature within the user granularity feature and the second product granularity feature within the product granularity feature; Based on the similarity, generate user importance features corresponding to the user granularity features and product importance features corresponding to the product granularity features; Read the preset latent semantic model matching formula, and determine the degree of matching based on the user granularity feature, the product granularity feature, the user importance feature, and the product importance feature according to the preset latent semantic model matching formula.

6. The method according to claim 5, characterized in that, The preset latent semantic model matching formula includes: in, Identify the user granularity features, Identify the particle size characteristics of the product. Identify the user's importance characteristics. Identify the important features of the product. , , The user's bias, the product to be matched, and the global bias are identified respectively.

7. A product matching device, characterized in that, The device includes: The information acquisition module is used to acquire users' historical review information and product introduction information of the products to be matched; The feature extraction module is used to determine user-level features in the historical review information and product-level features in the product introduction information according to a preset granularity. The product matching module is used to fuse the user granularity features and the product granularity features to determine the degree of matching between the product to be matched and the user; The preset granularity includes at least aspect granularity and comment granularity; The feature extraction module includes: A classification unit is used to divide the historical review information into at least one historical review information aspect set according to the product aspect to which it belongs, and to divide the product introduction information into at least one product introduction information set according to the product aspect to which it belongs; The comment feature unit is used to extract first-granularity user features corresponding to the comment granularity for the historical comment information and to extract first-granularity product features corresponding to the comment granularity for the product introduction information; An aspect feature unit is used to extract a second granular user feature corresponding to the aspect granularity for the historical review information aspect set and to extract a second granular product feature corresponding to the aspect granularity for the product introduction information set; the first granular user feature and the second granular user feature are used as the user granular feature, and the first granular product feature and the second granular product feature are used as the product granular feature; The product aspect refers to the aspects of analyzing and extracting user characteristics, including price, performance, service, and product type.

8. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor to enable the at least one processor to perform the product matching method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the product matching method of any one of claims 1-6.

Images