CN113127604B - Method and system for fine-grained item recommendation based on review text - Google Patents

Abstract

The present disclosure provides a method and system for fine-grained item recommendation based on comment text, including: obtaining a user comment text set and an item comment text set respectively; using a fine-grained feature interaction network to calculate a multi-granularity correlation matrix between users and items to obtain 3D interactive image; inputting the 3D interactive image into a fully connected neural network and a traditional factorization machine to achieve user-item scoring prediction, and implementing item recommendation according to the scoring result; the scheme is encoded with a multi-level dilated convolution structure User and item reviews avoid the loss of fine-grained information in reviews, construct feature interactions between user and item reviews at multiple granularities, and use 3D convolution to fuse multi-granularity information, effectively highlighting multiple granularities in reviews The relevant information below can effectively improve the rationality and accuracy of item recommendation.

Description

Method and system for fine-grained item recommendation based on review text

technical field

The present disclosure belongs to the technical field of item recommendation, and in particular, relates to a method and system for fine-grained item recommendation based on review text.

Background technique

The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

In many works, leveraging review texts to improve the performance of recommender systems has been shown to be effective. Review-based recommender systems can not only alleviate the cold-start problem, but also obtain more fine-grained representations of users and items. Earlier work mainly focused on topic modeling and language modeling methods. In recent years, researchers have used deep learning methods and achieved good results. ConvMF uses Convolutional Neural Networks (CNN) as automatic feature extractors to encode users (items) into low-dimensional vector representations. These are then combined using probabilistic matrix factorization (PMF) for scoring prediction. DeepCoNN (Deep Cooperative Neural Networks) uses two parallel convolutional architectures to infer latent feature representations from user and item reviews. Feature representations of users and items are concatenated and used as input to a factorization machine (FM) for rating prediction. Since the importance of different words for modeling is different, D-Attn introduces a word-level attention mechanism and assigns different weights to different words. DAML focuses on local information between users and item reviews and the interaction between the two. DAML integrates rating features and review features into a unified framework, and uses neural factorization machines to achieve high-order nonlinear interactions of features to complete the final rating prediction. Enhanced model interpretability. MPCN extracts important reviews from user and item reviews and then matches them verbatim. It not only makes predictions with the most informative reviews, but also enables deeper text interactions.

With the great success of deep learning in fields such as natural language processing, many recommendation models based on this technology have greatly improved the performance of rating prediction. In these works, convolutional neural network architectures extract latent features from user- and item-corresponding reviews, respectively. By using a dense word embedding matrix to represent reviews, a fixed-size sliding window is subsequently applied to capture contextual information. Convolutional neural network-based techniques better facilitate the understanding of semantic information in reviews, leading to significant improvements over existing bag-of-words model-based rating prediction methods. However, a much criticized shortcoming is that it is difficult for people to effectively understand the features extracted by neural networks, which limits the interpretability of recommender systems.

The key to review-based recommender systems lies in the capture of different granularity information in review texts. Generally speaking, the same user tends to have different preferences for different items, which is reflected in the reviews he writes as a whole. At the same time, important semantic features in reviews are implied in text fragments of different granularities. The example in the figure illustrates this problem. Different granularity of text (such as words, phrases, and sentences) can reveal user preferences for different items. Examples include words: "best", "great", "disappointing"; phrases: "quality great battery life", "screen looks amazing" and sentences: "I usually go with product red but it looked a bit orange this year.".

However, the inventors found that existing works usually learn a single feature vector for each user or item by integrating the review texts of users and items, and then achieve rating prediction by means such as factorization. These methods have limitations in capturing fine-grained user preferences and item attributes, as they only assign a single feature vector to user items, resulting in information at multiple granularities in reviews not being clearly captured.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present disclosure provides a method and system for fine-grained item recommendation based on review text. The solution provides a fine-grained feature interaction network that encodes user and item reviews with a multi-level dilated convolution structure. It avoids the loss of fine-grained information in reviews, constructs feature interactions between user and item reviews at multiple granularities, and uses 3D convolution to process multi-granularity information, effectively highlighting relevant information at multiple granularities in reviews. Effectively improve the rationality and accuracy of item recommendation.

According to a first aspect of the embodiments of the present disclosure, a fine-grained item recommendation method based on review text is provided, including:

Obtain the user comment text collection and the item review text collection respectively;

Use the fine-grained feature interaction network to calculate the multi-granularity correlation matrix between users and items to obtain 3D interactive images;

Inputting the 3D interactive image into a fully connected neural network and a traditional factorization machine to achieve user-item rating prediction, and implement item recommendation according to the rating result;

The fine-grained feature interaction network specifically includes: using a hierarchical dilated convolution model to extract multi-granularity semantic features for each comment text in the user comment text set and the item comment text set, respectively, to obtain multi-granularity comments on users and items. Represents a collection; computes a multi-granularity correlation matrix between users and items, and fuses the correlation matrix into a 3D interactive image through a 3D convolutional neural network.

Further, the use of the hierarchical dilated convolution model is different from the standard convolution to perform convolution on a continuous subsequence of the input at each step, and the dilated convolution of different levels skips δ input elements each time. The specific description is:

是向量拼接，b是偏置项，ReLU是非线性激活函数，x_t表示上下文中心单词，W为大小为2w+1的卷积核,w为用于调整卷积核大小的参数，δ为扩张率，其为不小于1的整数。in,

is the vector splicing, b is the bias term, ReLU is the nonlinear activation function, x _t represents the context center word, W is the convolution kernel of size 2w+1, w is the parameter used to adjust the size of the convolution kernel, and δ is the expansion rate, which is an integer not less than 1.

Further, the calculation of the multi-granularity correlation matrix between the user and the item specifically includes: realizing the interaction of the features between the user and the item by constructing the correlation matrix on each granularity, given the nth and the th of the user and the item. There are m multi-granularity comment representations, then the relationship matrix of the l-th granularity is realized by the following dot product operation:

中的元素

代表着用户与物品评论中第i个特征与第j个特征在l粒度下的相关性。where l=[1,2,...,L] is the number of semantic granularities,

elements in

Represents the correlation between the ith feature and the jth feature in user and item reviews at l granularity.

According to a second aspect of the embodiments of the present disclosure, a fine-grained item recommendation system based on review text is provided, including:

a data acquisition unit, which is used to acquire the user comment text set and the item comment text set respectively;

A multi-granularity feature extraction unit, which is used to extract multi-granularity semantic features for each comment text in the user comment text set and the item comment text set respectively by using a hierarchical dilated convolution model to obtain a multi-granularity comment representation set of users and items;

an interactive image generation unit, which is used to construct a multi-granularity correlation matrix between users and items based on the multi-granularity comment representation of users and items, and fuse the correlation matrix into a 3D interactive image through a 3D convolutional neural network;

An item recommendation unit, which is used to input the 3D interactive image into a fully connected neural network and a traditional factorization machine to achieve user-item rating prediction, and implement item recommendation according to the rating result.

According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, the processor implements the comment-based method when the processor executes the program A fine-grained item recommendation system for text.

According to a fourth aspect of the embodiments of the present disclosure, there is provided a non-transitory computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the described fine-grained comment text-based Item recommendation system.

Compared with the prior art, the beneficial effects of the present disclosure are:

(1) The solution of the present disclosure provides a fine-grained feature interaction network, and the multi-granularity correlation matrix between users and items is calculated through the fine-grained feature interaction network. The advantages are: multi-level user/item representation Interacting with fine-grained features, the network does not represent each user with a single abstract vector, but uses hierarchical dilated convolutions in a unified module to build comment-based multi-level representations; In convolution, the received input width of each layer grows exponentially, while the number of parameters increases only linearly. At the same time, the output of each layer is preserved as feature maps across text segments of different lengths, without any coverage loss due to the absence of any form of pooling or stepped convolution; in this way, the scheme can Semantic features in reviews are gradually obtained at different granularities through local correlations and long-term dependencies, including word, phrase, and sentence levels. The proposed scheme encodes user and item reviews with a multi-level dilated convolution structure, avoiding the loss of fine-grained information in reviews.

(2) In order to avoid information loss, the fine-grained feature interaction network constructs feature interactions between user reviews and item reviews at each semantic granularity. In practical applications, the model is based on the hierarchical representation method of comment text, and builds an association matrix from word to sentence level for each pair of comments. In this way, multiple granular features implicit in user and item reviews can be identified and interacted with with minimal loss, providing sufficient contextual relevance cues for predicting accurate ratings. Then, we merge the multiple association matrices of review pairs at each granularity into a 3D image whose channels represent the correlation degree of user and item review feature interactions at different granularities. Predict user ratings for items by identifying high-order salient signals through a hierarchy similar to 3D convolution-based image recognition; the proposed scheme builds feature interactions between users and item reviews at multiple granularities, and uses 3D convolution to fuse multiple Granular information, effectively highlighting the relevant information under multiple granularities in the comments.

(3) The proposed scheme is not only superior in performance, but also has strong interpretability.

Advantages of additional aspects of the disclosure will be set forth in part in the description that follows, and in part will become apparent from the description below, or will be learned by practice of the disclosure.

Description of drawings

The accompanying drawings that constitute a part of the present disclosure are used to provide further understanding of the present disclosure, and the exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation of the present disclosure.

1 is a schematic structural diagram of the fine-grained feature interaction network model described in Embodiment 1 of the present disclosure;

FIG. 2 is the hierarchical dilated convolution described in Embodiment 1 of the present disclosure.

Detailed ways

The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the present disclosure. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates that There are features, steps, operations, devices, components and/or combinations thereof.

The embodiments of this disclosure and features of the embodiments may be combined with each other without conflict.

Example 1:

The purpose of this embodiment is to provide a fine-grained item recommendation method based on review text.

A fine-grained item recommendation method based on review text, including:

Obtain the user comment text collection and the item review text collection respectively;

Use the fine-grained feature interaction network to calculate the multi-granularity correlation matrix between users and items to obtain 3D interactive images;

Inputting the 3D interactive image into a fully connected neural network and a traditional factorization machine to achieve user-item rating prediction, and implement item recommendation according to the rating result;

The fine-grained feature interaction network specifically includes: using a hierarchical dilated convolution model to extract multi-granularity semantic features for each comment text in the user comment text set and the item comment text set, respectively, to obtain multi-granularity comments on users and items. Represents a collection; computes a multi-granularity correlation matrix between users and items, and fuses the correlation matrix into a 3D interactive image through a 3D convolutional neural network.

Specifically, in order to facilitate understanding, the solutions described in the present disclosure are described in detail below with reference to the accompanying drawings:

Figure 1 shows a schematic diagram of the model structure of the fine-grained feature interaction network, which in turn describes the fine-grained feature interaction network (FFIN: fine-grained feature interaction network) proposed by us. First, we define the review-based recommendation problem; then, we describe in detail the encoding process of the original review text by the hierarchical dilated convolutional structure in the review representation module; secondly, we show the multi-granular review representation of users and items feature interaction process; finally, we introduce a rating prediction layer with a factorization machine (FM) to achieve user ratings for items, where:

(1) Definition of the problem

和

其中N，M对应用户和物品评论集合的评论数量，其中，用户评论文本集合，其存储的是特定用户关于不同物品的评论文本；物品评论文本集合，其存储的是关于特定物品不同用户的评论文本。我们的目的是建立一个预测模型，通过一组用户和物品的评论集合来预测用户对该物品的评分。设一组用户和物品评论集合中单个评论的最大长度分别为Tu和Ti，则第n个用户和第m个物品原始的评论文本可以分别表示为：The review-based recommendation problem can be formulated as: for a given user u and item i, they both correspond to a set of original review texts

and

Among them, N and M correspond to the number of comments in the user and item comment sets, among which, the user comment text set stores the comment text of a specific user about different items; the item comment text set stores the comments of different users about a specific item text. Our goal is to build a predictive model that predicts a user's rating for an item from a set of user and item reviews. Assuming that the maximum lengths of a single comment in a set of user and item comments are Tu and Ti, respectively, the original comment texts of the nth user and the mth item can be expressed as:

(2) Representation of comment text

简化为s，评论长度设为T。Since the generation process of multi-granular review representations for users and items is similar, we only show the encoding steps for a single review text, and

Simplify to s, and the comment length is set to T.

Inspired by the multi-layer dilated convolution structure, as shown in Figure 2, we design a hierarchical dilated convolution (HDC) to learn multiple semantic-grained review representations directly from reviews.

For a given review text s=[w ₁ , w ₂ , ..., w _T ], the model first maps it to the corresponding word vector matrix through an embedding matrix, namely X=[x ₁ , x ₂ , . .., x _T ], where x _t ∈ R ^d is the word embedding representation of the t-th word in this text, and d is the dimension size of the word vector. Then, HDC takes this word vector matrix as input to capture the multi-granularity semantic features in the reviews.

Unlike standard convolution which convolves a contiguous subsequence of the input at each step, dilated convolution has a wider receptive field by skipping δ input elements at a time, where δ is the dilation rate. For a context center word _xt and a convolution kernel W of size 2w+1, the dilated convolution operation can be formulated as:

是向量拼接、b是偏置项、ReLU是非线性激活函数。如图2所示，每个卷积层的输出是前一层输入的加权组合。我们以δ＝1为开始(等同于标准卷积)，以确保不遗漏原始输入序列中的任何元素。之后，通过以更大的扩张率层次叠加扩张之后的卷积，卷积文本的长度以指数形式扩展，只需使用少量的层和适量的参数就可以覆盖不同接受域的语义特征。in,

is the vector concatenation, b is the bias term, and ReLU is the nonlinear activation function. As shown in Figure 2, the output of each convolutional layer is a weighted combination of the inputs of the previous layer. We start with δ=1 (equivalent to standard convolution) to ensure that no elements in the original input sequence are missed. Afterwards, by stacking the dilated convolutions hierarchically with a larger dilation rate, the length of the convolutional text expands exponentially, covering the semantic features of different receptive fields with only a small number of layers and appropriate parameters.

其中f是是每一层卷积滤波器的数量。假设有L层堆叠的扩张卷积，多粒度评论表示可以定义为[s¹，s²，...，s^L]。通过这种方式，HDC以较小的扩张率从单词和短语层次逐步收获词义和语义特征，以较大的扩张率从句子层次捕获长期的依赖关系。我们的评论编码模块不仅在并行能力上优于循环神经网络，而且相比于完全基于注意力的方法能够显著减少内存的消耗。Furthermore, to prevent vanishing or exploding gradients, we apply layer normalization at the end of each convolutional layer. Since there may be irrelevant information introduced into long-distance semantic units, we actually design multi-level dilation rates according to the performance in validation. The output of each stacked layer l is saved as a feature map of the text at a specific granularity level, the formula is

where f is the number of convolutional filters in each layer. Assuming there are L layers of stacked dilated convolutions, the multi-granularity comment representation can be defined as [s ¹ , s ² , ..., s ^L ]. In this way, HDC gradually harvests word sense and semantic features from the word and phrase levels with a small dilation rate, and captures long-term dependencies from the sentence level with a large dilation rate. Our comment encoding module not only outperforms recurrent neural networks in parallelism, but also significantly reduces memory consumption compared to fully attention-based methods.

After the HDC module encoding, the multi-granularity comment representation of a single user and item can be expressed as:

分别为用户和物品评论集合S_u、S_i中的第n个和第m个评论文本。in,

are the nth and _mth review texts in the user and item review sets _Su , Si, respectively.

(3) Feature interaction module

和

我们通过在每个粒度上构建关联矩阵实现用户和物品之间特征的交互。给定用户和物品的第n个和第m个多粒度评论表示，第l层粒度的关联矩阵可以通过如下的点乘运算实现：A collection of multi-granular review representations for a given user and item

and

We implement the interaction of features between users and items by building an association matrix at each granularity. Given the n-th and m-th multi-granularity review representations of users and items, the relationship matrix of the l-th granularity can be realized by the following dot product operation:

中的元素

代表着用户与物品评论中第i个特征与第j个特征在l粒度下的相关性。where l=[1,2,...,L] is the number of semantic granularities,

elements in

Represents the correlation between the ith feature and the jth feature in user and item reviews at l granularity.

In order to summarize the multi-granularity features in the reviews of the entire user and item, FFIN fuses the correlation matrix obtained from all interactions between the user and the item into a 3D interaction image, the formula is:

where n×m represents the total number of correlation matrices under a single granularity, and each pixel in the 3D interactive image is obtained in the following way:

Specifically, each pixel is a spliced vector of features at all granularities, representing the degree of interaction between users and items at multiple granularities.

Since the user's rating behavior is usually subjective and personalized, it will reflect different preferences when faced with different items. We employ a layered 3D convolutional neural network and a max-pooling operation to identify salient matching signals from the entire image. The 3D convolution is an extension of the typical 2D convolution, and its filter and stride are 3D cubes. Formally, the high-order pixels on the z-th feature map of the t-th layer are computed at (k, i, j) as:

和b^(t)分别是3D卷积核和偏置项，大小为W_t×H_t×R_t。随后，通过一个最大池化操作提取出显著的信息：Among them, ELU is a nonlinear activation function, z' represents the feature map of the previous layer,

and b ^(t) are the 3D convolution kernel and bias term, respectively, of size W _t ×H _t ×R _t . Subsequently, significant information is extracted by a max-pooling operation:

和

是3D池化操作的大小。最后一层的输出是将用户和物品评论之间的综合交互向量进行拼接，表示为

in,

and

is the size of the 3D pooling operation. The output of the last layer is to concatenate the comprehensive interaction vectors between users and item reviews, expressed as

(4) Score prediction module

输入到全连接神经网络和传统的因子分解机(FM)中来实现最终的用户-物品评分预测。全连接神经网络的实现过程可以被公式如下：In this subsection, we combine the prediction vectors produced by the feature interaction module above

The inputs are fed into a fully connected neural network and a traditional Factorization Machine (FM) to achieve the final user-item rating prediction. The implementation process of the fully connected neural network can be formulated as follows:

Among them, W ₀ ∈ R ^n×e and b ₀ ∈ R ⁿ are the weight parameters and bias terms of the fully connected network, respectively.

FM takes a real-valued feature vector and uses decomposition parameters to model the interactions between features. Its definition is as follows

是实值输入特征向量。<.，.>是点乘操作。参数{v₁，...，v_n}，υ∈Rⁿ是用于建模交互对(x_i，x_i)的分解参数。

和

分别是全局偏置项和线性回归组件。FM的权重参数。FM的输出

是一个标量，代表用户和物品交互的预测评分。in,

is the real-valued input feature vector. <., .> is the dot product operation. The parameters {v ₁ , . . . , v _n }, υ∈R ⁿ are the decomposition parameters used to model the interaction pair ( _xi , x _i ).

and

are the global bias term and the linear regression component, respectively. FM weight parameter. FM output

is a scalar representing the predicted rating of the user-item interaction.

Further, different items are sorted in descending order according to their predicted scores, and items with high predicted scores are recommended to users.

Embodiment 2:

The purpose of this embodiment is to provide a fine-grained item recommendation system based on comment text.

A fine-grained item recommendation system based on review text, including:

a data acquisition unit, which is used to acquire the user comment text set and the item comment text set respectively;

A multi-granularity feature extraction unit, which is used to extract multi-granularity semantic features for each comment text in the user comment text set and the item comment text set respectively by using a hierarchical dilated convolution model to obtain a multi-granularity comment representation set of users and items;

an interactive image generation unit, which is used to construct a multi-granularity correlation matrix between users and items based on the multi-granularity comment representation of users and items, and fuse the correlation matrix into a 3D interactive image through a 3D convolutional neural network;

An item recommendation unit, which is used to input the 3D interactive image into a fully connected neural network and a traditional factorization machine to achieve user-item rating prediction, and implement item recommendation according to the rating result.

In further embodiments, there is also provided:

An electronic device includes a memory, a processor, and computer instructions stored on the memory and executed on the processor, and when the computer instructions are executed by the processor, the method described in the first embodiment is completed. For brevity, details are not repeated here.

It should be understood that, in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate array FPGA or other programmable logic devices , discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory may include read-only memory and random access memory and provide instructions and data to the processor, and a portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in the first embodiment is completed.

The method in the first embodiment can be directly embodied as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The software modules may be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other storage media mature in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and completes the steps of the above method in combination with its hardware. To avoid repetition, detailed description is omitted here.

Those of ordinary skill in the art can realize that the unit, that is, the algorithm step of each example described in conjunction with this embodiment, can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

The method and system for fine-grained item recommendation based on review text provided by the above embodiments can be implemented and have broad application prospects.

The above descriptions are only preferred embodiments of the present disclosure, and are not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they do not limit the protection scope of the present disclosure. Those skilled in the art should understand that on the basis of the technical solutions of the present disclosure, those skilled in the art do not need to pay creative efforts. Various modifications or variations that can be made are still within the protection scope of the present disclosure.

Claims (10)

1. A fine-grained item recommendation method based on comment texts is characterized by comprising the following steps:

respectively acquiring a user comment text set and an article comment text set; the user comment text set and the article comment text set respectively correspond to a group of comment texts;

calculating a multi-granularity incidence matrix between a user and an article by using a fine-granularity feature interaction network to obtain a 3D interaction image;

inputting the 3D interactive image into a fully-connected neural network and a traditional factorization machine to realize the grading prediction of a user-article, and realizing the recommendation of the article according to the grading result;

the fine-grained feature interaction network specifically includes: performing multi-granularity semantic feature extraction on each comment text in the user comment text set and the article comment text set respectively by using a hierarchical expansion convolution model to obtain multi-granularity comment expression sets of the user and the article; and calculating a multi-granularity incidence matrix between the user and the object, and fusing the incidence matrix into a 3D interactive image through a 3D convolutional neural network.

2. A comment text based fine grain item recommendation method according to claim 1, wherein the use of the hierarchical dilation convolution model to convolve a continuous subsequence of inputs at each step differently from the standard convolution, the different levels of dilation convolution operate by skipping δ input elements at a time, and the dilation operation is specifically described as:

wherein,

is vector concatenation, b is a bias term, reLU is a non-linear activation function, x _t Representing the context center word, W is a convolution kernel of size 2w +1, W is a parameter for adjusting the size of the convolution kernel, and δ is an expansion rate, which is an integer not less than 1.

3. A comment text based fine grain item recommendation method as claimed in claim 2 wherein to prevent the disappearance or explosion of the gradient, the last application layer of the dilation convolution at each layer is normalized.

4. The method for recommending fine-grained items based on comment texts as claimed in claim 1, wherein said calculating a multi-grained association matrix between the user and the item specifically comprises: the interaction of features between users and articles is realized by constructing a correlation matrix on each granularity, and given the nth and mth multi-granularity comment representations of the users and the articles, the correlation matrix of the ith layer of granularity is realized by the following dot product operation:

wherein L = [1,2]As to the number of semantic granularities,

element (1) of

Representing the correlation between the ith feature and the jth feature in the user and item reviews under the granularity of l, and f is the number of convolution filters in each layer.

5. The fine-grained item recommendation method based on comment text according to claim 1, wherein the correlation matrix is fused into a 3D interactive image through a 3D convolutional neural network, wherein each pixel point in the 3D interactive image is a spliced vector of features at all granularities, and represents the degree of interaction between a user and an item at multiple granularities.

6. The comment text-based fine-grained item recommendation method according to claim 1, wherein the 3D interactive image is input into a fully-connected neural network and a traditional factorization machine to realize user-item rating prediction, and specifically comprises the following steps: inputting the 3D interactive image into a full-connection neural network to obtain a real-value feature vector, inputting the feature vector into a factor decomposition machine, modeling the interaction between features by using decomposition parameters, and outputting a prediction score of the interaction between the user and the article.

7. The comment text-based fine-grained item recommendation method according to claim 1, wherein before multi-grained semantic feature extraction is performed on comment texts by adopting a hierarchical expansion convolution model, each comment text is mapped into a corresponding word vector matrix through an embedding matrix.

8. A comment text-based fine-grained item recommendation system is characterized by comprising:

the data acquisition unit is used for respectively acquiring a user comment text set and an article comment text set; the user comment text set and the article comment text set respectively correspond to a group of comment texts;

the multi-granularity feature extraction unit is used for respectively extracting multi-granularity semantic features of each comment text in the user comment text set and the article comment text set by utilizing the hierarchical expansion convolution model to obtain multi-granularity comment expression sets of the user and the article;

the interactive image generation unit is used for constructing a multi-granularity incidence matrix between the user and the article based on the multi-granularity comment expression of the user and the article, and fusing the incidence matrix into a 3D interactive image through a 3D convolutional neural network;

and the item recommending unit is used for inputting the 3D interactive image into the fully-connected neural network and the traditional factorization machine to realize the grading prediction of the user-item and realize the recommendation of the item according to the grading result.

9. An electronic device comprising a memory, a processor, and a computer program stored and executed on the memory, wherein the processor when executing the program implements a comment text based fine grain item recommendation system as recited in any one of claims 1 to 7.

10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor implements a comment text based fine grain item recommendation system as recited in any one of claims 1 to 7.

Images