Robust recommendation method and system for reducing the proportion of useless information in attack environment

Through the combination of two-way LSTM and attention mechanism, the problem of useless information filtering of recommendation system in the attack environment is solved, the accuracy and robustness of recommendations are improved, and the authenticity and interpretability of information are ensured.

CN114385926BActive Publication Date: 2024-11-15SHANDONG NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111462560.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-02
Publication Date
2024-11-15
Estimated Expiration
2041-12-02

AI Technical Summary

Technical Problem

Existing recommendation systems are vulnerable to false information attacks in attack environments, resulting in a decrease in recommendation accuracy and lack of effective robustness and interpretability, especially the filtering of useless information is difficult to achieve.

Method used

Two-way LSTM is used to learn emotional polarity marking and attention mechanisms, filter out useful comment information, reduce the proportion of useless information, and improve recommendation accuracy through predictive recommendation models.

Benefits of technology

It effectively reduces the proportion of useless information, improves the robustness and accuracy of the recommendation system, enhances the interpretability of information, and ensures the authenticity of feature learning information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114385926B_ABST
    Figure CN114385926B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of recommendation system and robustness research, and provides a robust recommendation method and system for reducing the proportion of useless information in an attack environment. The method includes obtaining the comment information in each item; marking the sentiment polarity of the comment information, comparing it with the corresponding rating score after marking, and marking it as useless information if the difference in sentiment polarity between the two exceeds a set threshold, otherwise, marking it as real information; based on the marked useless information and real information, using a predictive recommendation model to obtain the feature representation of each user or project; using an attention mechanism to learn the weight of each comment, respectively combining the feature representation of the user or project, to obtain a recommendation result that reduces the proportion of useless users.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of recommendation system and robustness research, and in particular relates to a robust recommendation method and system for reducing the proportion of useless information in an attack environment. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the advancement of science and technology, electronic information has penetrated into all areas of life, but it has also brought about information explosion and information overload. The emergence of recommendation systems has solved this problem very well. The main idea of ​​the recommendation system is to predict the user's preference for items based on the user's historical behavior, so as to recommend effective information to the user and improve the user's search efficiency. Collaborative filtering is an advanced recommendation method that models user preferences and item features by combining different historical records such as comments, ratings, and historical browsing. It can be divided into memory-based and model-based methods. Although collaborative filtering has good performance in many recommendation scenarios, it encounters an important problem in some practical applications: it is not explainable when recommending certain items. The lack of explainable recommendations reduces the user's trust and the degree of research by subsequent researchers.

[0004] In most recommendation systems, users can freely edit text reviews and rating information. Text reviews contain rich information about product items and user experience, and sometimes also provide guidance. However, due to the subjective differences in user reviews, not all reviews are valid, and it is difficult to obtain valid information from a large number of reviews.

[0005] Many studies have shown that the economic benefits brought by the widespread use of recommendation systems have attracted a group of malicious attackers to disrupt the system. Attackers intentionally increase or decrease the ranking of certain items by publishing false information in the recommendation system, thereby increasing or decreasing the recommendation degree and ranking of these attacked items. This behavior of destroying the recommendation system is called a shilling attack, which seriously disrupts the fairness of the recommendation system. How to alleviate the decline in recommendation accuracy caused by the attack problem of the recommendation system has been a hot topic in the robustness research of the recommendation system in recent years. At present, there are three main ways to prevent and control attacks on recommendation systems: one is to enhance the robustness of the recommendation algorithm under the premise of known attack behaviors, that is, the robustness or accuracy of the recommendation performance when attacked. The second is to efficiently detect attacking users, malicious ratings and comments, and false profiles. The third is to eliminate attacking users from the source, which requires the system to have a more complete user screening system mechanism. Through research, the performance of attacks is diverse. Some attackers disrupt the reference value or recommendation ranking of certain items by publishing useless information. The useless information is mainly manifested in the contradiction of the emotional polarity of ratings and comments.

[0006] In recent years, deep learning has made significant research breakthroughs in fields such as natural language processing. Some recommendation systems have tried to combine different neural networks with collaborative filtering to improve recommendation performance and accuracy. In the earliest stages, feature extraction and sentiment analysis of comment information usually require manual preprocessing. For text sentiment analysis techniques such as EFM, it is a simple extraction of words or phrases from the text, which may distort the meaning or depth of the text. The present invention uses a bidirectional LSTM to perform sentiment polarity analysis and an end-to-end neural network to select useful comments and enhance information interpretability. Summary of the invention

[0007] In order to solve the technical problems existing in the above-mentioned background technology, the present invention provides a robust recommendation method and system for reducing the proportion of useless information in an attack environment. The data set preprocessing marks the useless information injected by the attacking user, reduces the proportion of useless information in the subsequent recommendation model to ensure the authenticity of the subsequent feature learning information source, and uses the attention mechanism to screen out comment information with reference value for learning to improve the recommendation accuracy.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A first aspect of the present invention provides a robust recommendation method for reducing the proportion of useless information in an attack environment.

[0010] Robust recommendation methods to reduce the proportion of useless information in an attack environment include:

[0011] Get the comment information in each entry;

[0012] The comment information is marked with sentiment polarity, and after marking, it is compared with the corresponding rating score. If the difference in sentiment polarity between the two exceeds a set threshold, it is marked as useless information, otherwise, it is marked as true information;

[0013] Based on the marked useless information and real information, a predictive recommendation model is used to obtain the feature representation of each user or project. The attention mechanism is used to learn the weight of each comment, combined with the feature representation of each user or project, to obtain a recommendation result that reduces the proportion of useless users.

[0014] A second aspect of the present invention provides a robust recommendation system that reduces the proportion of useless information in an attack environment.

[0015] A robust recommendation system that reduces the proportion of useless information in an attack environment, including:

[0016] An acquisition module is configured to: acquire comment information in each entry;

[0017] The sentiment analysis module is configured to: mark the sentiment polarity of the comment information, compare it with the corresponding rating score after marking, and if the difference in sentiment polarity between the two exceeds a set threshold, mark it as useless information, otherwise, mark it as true information;

[0018] The prediction and recommendation module is configured as follows: based on the marked useless information and real information, a prediction and recommendation model is used to obtain the feature representation of each user or project; an attention mechanism is used to learn the weight of each comment, and combined with the feature representation of each user or project, a recommendation result that reduces the proportion of useless users is obtained.

[0019] A third aspect of the present invention provides a computer-readable storage medium.

[0020] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the robustness recommendation method for reducing the proportion of useless information in an attack environment as described in the first aspect above.

[0021] A fourth aspect of the present invention provides a computer device.

[0022] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the robustness recommendation method for reducing the proportion of useless information in an attack environment as described in the first aspect above are implemented.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] In the useless information search part of the present invention, we use a two-layer LSTM model to mark the sentiment polarity of user comment texts, and compare them with the corresponding rating information after marking. If the difference in sentiment polarity between the two exceeds the set threshold, it is marked as useless information.

[0025] The present invention introduces a neural mechanism to establish a recommendation model, models and learns users and projects respectively, and introduces an attention mechanism to evaluate the usefulness of comments.

[0026] The present invention ultimately uses a neural model to use user preferences and project features to predict ratings, and ultimately outputs a recommendation result with a higher accuracy that reduces the proportion of useless users. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0028] Figure 1is a flow chart of a robust recommendation method for reducing the proportion of useless information in an attack environment according to an embodiment of the present invention;

[0029] Figure 2 This is a working framework diagram of data sentiment analysis preprocessing and marking useless users in an embodiment of the present invention;

[0030] Figure 3 It is a detailed framework diagram of the embodiment of the present invention for learning user and project features to perform rating prediction and recommendation. DETAILED DESCRIPTION

[0031] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0033] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0034] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the methods and systems according to various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, program segment, or a part of a code may include one or more executable instructions for implementing the logical functions specified in each embodiment. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and / or block diagram, and the combination of boxes in the flowchart and / or block diagram can be implemented using a dedicated hardware-based system that performs a specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

[0035] Embodiment 1

[0036] like Figure 1As shown, this embodiment provides a robust recommendation method for reducing the proportion of useless information in an attack environment. This embodiment uses the method applied to a server as an example. It can be understood that the method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communications, middleware services, domain name services, security services CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, which is not limited in this application. In this embodiment, the method includes the following steps:

[0037] S101: Obtain comment information in each entry;

[0038] Specifically, the recommendation data set usually contains metadata information such as ratings, comments, project IDs, user IDs, etc. for each item. This embodiment extracts the comment information of each item. The components for implementing this step are bidirectional LSTM and Glove (Glove is a set of pre-trained word vectors in the field of sentiment analysis). First, the word segmentation and word embedding of each comment information are performed, and the dictionary is used for storage and comparison, and finally the positive (positive) or negative (negative) result of the predicted text is generated. Bidirectional LSTM has the characteristics of combining positive and negative short-term memory, and uses the loss function sigmoid to train its loss. Compared with other recurrent neural networks, the final prediction of the sentiment polarity result reaches a certain higher accuracy.

[0039] S102: marking the sentiment polarity of the comment information, and comparing it with the corresponding rating score after marking. If the difference in sentiment polarity between the two exceeds a set threshold, it is marked as useless information, otherwise, it is marked as real information; Figure 2 shown.

[0040] The predicted sentiment polarity of each comment is compared with the rating result corresponding to the comment to determine and mark useless information. Set a threshold. In the scoring mechanism in most recommendation fields, the rating score is a score in the interval [1, 5], with 1 being the lowest score for a bad review and 5 being the highest score for a good review. If the difference between the sentiment polarity of a comment and the rating score ([1, 5]) is lower or higher than the threshold, it is judged as useful information or useless information. Specifically, the middle value 3 is set as the sentiment polarity dividing line within the rating score interval. Ratings greater than 3 are considered positive sentiments, and ratings less than 3 are considered negative sentiments. Compare the rating sentiment polarity with the comment sentiment polarity. If the two are inconsistent, mark it as useless information.

[0041] S103: Based on the marked useless information and real information, a prediction recommendation model is used to obtain a feature representation of each user or item;

[0042] All real information and some useless information are injected into the predictive recommendation model. For users and projects respectively, each comment is converted into a word vector matrix, and the matrix is ​​input into the text processor of the CNN convolutional neural network. The first layer of the CNN convolutional neural network text processor is the embedding layer, which maps each word in the comment matrix into a d-dimensional vector, and then converts it into an embedding matrix of fixed length. The embedding can be a pre-trained embedding. The embedding layer embeds the vocabulary information and then inputs it to the convolution layer.

[0043] Convolution operation generates feature representation. The convolution layer consists of multiple neurons and filters, and features are generated through convolution operation. In the final output, all feature vectors of a user or project are obtained, and all vector representations of a user or project are aggregated to obtain the final representation of the project or user.

[0044] S104: Use the attention mechanism to learn the weight of each comment, and combine it with the feature representation of a user or project to obtain a recommendation result that reduces the proportion of useless users. Figure 3 shown.

[0045] The attention mechanism learns the weight of comments. Useless users are marked and their weight is reduced. It is assumed that the importance of each comment in the remaining information to the project is consistent, but this is not the case in real life. Some users' comments may be perfunctory, and the comments they edit may have no reference value. In order to alleviate the deviation caused by this problem on the recommendation performance, this embodiment uses the attention mechanism to learn to evaluate the weight of each comment to a certain extent, select comments that are meaningful to the project, and then aggregate these useful information to describe the project or user. In order to identify users who output meaningless comments, this embodiment embeds the unique information of each user and project, namely the ID information, to simulate user quality, and improve the generalization effect of the model with targeted information, and finally outputs a discrete attention score. In order to normalize the data to a certain limited interval, the softmax function is used to standardize the normalized attention score, and the feature vectors of the project and the user are the weighted sum of the features of all comments. Finally, the project and user vectors are sent to a fully connected layer to obtain the final representation.

[0046] This embodiment uses the LFM component to complete the prediction rating and generate the final recommendation. Specifically, the latent representation learned from the user and the project is introduced, and the user representation and the project representation are added to each other and then the two parts of information are dot-multiplied to output an n-dimensional vector, which is then passed to the prediction layer. The prediction layer multiplies the weight with the n-dimensional vector, and then adds it to the user, project and global deviation to obtain the predicted rating score.

[0047] Learning training model. The final result of this embodiment is the rating prediction value of the information that reduces the proportion of useless users, that is, the rating prediction score of all items. It is actually a regression problem, and the square loss function is often used for training and learning. The formula is as follows:

[0048]

[0049] in, is the predicted rating value, and R u,i is the actual rating value.

[0050] As one or more implementation modes, the process of obtaining the feature representation of each user or item includes:

[0051] Input the labeled useless information and real information into the prediction and recommendation model to obtain all feature vectors of a certain user or item;

[0052] Aggregate all feature vectors of a user or item to get the feature representation of the item or user.

[0053] As one or more implementation modes, the process of learning the weight of each comment using the attention mechanism includes:

[0054] The attention mechanism is used to learn the weight of each comment, combined with ID embedding to simulate the quality of projects and users, and the softmax function is used to obtain the standardized attention score.

[0055] As one or more implementations, the process of combining the feature representation of each user or item includes:

[0056] Multiply the feature representation of each user or item by the weight of all comments to obtain the feature vector of a user or item.

[0057] As one or more implementation modes, the process of obtaining a recommendation result for reducing the proportion of useless users includes:

[0058] Based on the feature vectors of all users or items, predict the recommendation results that reduce the proportion of useless users.

[0059] In the process of verifying the effectiveness of this embodiment, this embodiment uses two datasets: the Yelp2017 dataset and the Toys and Game dataset. Among them, Toys and Game is a smaller dataset, which only contains about 160,000 reviews. Yelp2017 is a larger dataset containing metadata information such as restaurant ratings and reviews. Although the dataset contains about 3 million reviews, it still has a certain sparsity. Previous studies have shown that sparsity will have a certain impact on the accuracy of recommendations, so the dataset is preprocessed. After the preprocessing, each user and project in the dataset contains at least 5 reviews. The rating interval of the two datasets is [1,5], and the rating increases in sequence according to the user's preference for the project, and the dataset also contains other metadata information.

[0060] The evaluation indicators used to verify this embodiment are the root mean square error (Root Mean Square Error) and the mean absolute error (Mean Absolute Error). The smaller the evaluation indicator value, the better the rating prediction performance of the project, and the smaller the difference between the predicted value and the true value, so that the accuracy of recommendation to the user is improved. The calculation method is as follows:

[0061]

[0062]

[0063] This embodiment first extracts the comment information of each item in the data set, uses the sentiment analysis model to analyze the sentiment polarity of the comment information and compares it with the corresponding rating information to complete the labeling of useless users. In the recommendation prediction model, the user and project are respectively modeled to extract features, and a prediction component is added to the top of the two networks to combine the user and project feature representations for rating prediction. The sentiment analysis model composed of bidirectional LSTM and Glove and used to identify and annotate useless information in the data set and the prediction recommendation model that learns users and projects separately to generate project prediction scores are trained to achieve a certain accuracy rate, and finally present a more realistic and accurate recommendation result to the user.

[0064] Embodiment 2

[0065] This embodiment provides a robust recommendation system that reduces the proportion of useless information in an attack environment.

[0066] A robust recommendation system that reduces the proportion of useless information in an attack environment, including:

[0067] An acquisition module is configured to: acquire comment information in each entry;

[0068] The sentiment analysis module is configured to: mark the sentiment polarity of the comment information, compare it with the corresponding rating score after marking, and if the difference in sentiment polarity between the two exceeds a set threshold, mark it as useless information, otherwise, mark it as true information;

[0069] The prediction and recommendation module is configured as follows: based on the marked useless information and real information, a prediction and recommendation model is used to obtain the feature representation of each user or project respectively; an attention mechanism is used to learn the weight of each comment, and the user or project is combined to generate a feature representation, so as to obtain a recommendation result that reduces the proportion of useless users.

[0070] It should be noted that the examples and application scenarios implemented by the acquisition module, sentiment analysis module and prediction and recommendation module are the same as those in the steps of Embodiment 1, but are not limited to the contents disclosed in Embodiment 1. It should be noted that the above modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0071] Embodiment 3

[0072] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the robustness recommendation method for reducing the proportion of useless information in an attack environment as described in the first embodiment above are implemented.

[0073] Embodiment 4

[0074] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the robustness recommendation method for reducing the proportion of useless information in an attack environment as described in the first embodiment above are implemented.

[0075] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program code.

[0076] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0077] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0078] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0079] A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A robust recommendation method for reducing the proportion of useless information in an attack environment, characterized in that: include: Obtaining review information in each entry, wherein each entry includes a rating score, review information, item ID, and user ID; Extract the comment information of each entry, including: segmenting and embedding each comment information through bidirectional LSTM and Glove, storing and comparing using a dictionary, and generating positive or negative results for predicting the text; The emotional polarity of the comment information is predicted to obtain the predicted emotional polarity of each comment; the predicted emotional polarity of each comment is compared with the rating result corresponding to the comment, and useless information is determined and marked, including: in the scoring mechanism in the recommendation field, the rating score is a score in the interval [1, 5], 1 is the lowest score for a bad review, 5 is the highest score for a good review, and the middle value 3 is set in the interval of the rating score as the emotional polarity dividing line, a rating greater than 3 is regarded as a positive emotion, and a rating less than 3 is regarded as a negative emotion; the rating emotional polarity is compared with the comment emotional polarity, and if the difference between the emotional polarities of the two exceeds a set threshold, it is marked as useless information, otherwise, it is marked as real information; Based on the labeled useless information and real information, a predictive recommendation model is used to obtain the feature representation of each user or item; The attention mechanism is used to learn the weight of each comment, including: using the attention mechanism to learn and evaluate the weight of each comment, selecting comments that are meaningful to the project, and then aggregating useful information to describe the project or user; in order to identify users who output meaningless comments, the unique information of each user and project, namely ID information embedding, is introduced to simulate user quality, and targeted information is used to improve the generalization effect of the model, and finally discrete attention scores are output, and the softmax function is used to standardize the discrete attention scores; Combining the feature representation of each user or item respectively, including: multiplying the feature representation of each user or item by the weight of all comments respectively, and then obtaining the feature vector of the user or item respectively; Use the LFM component to complete the prediction rating and generate the final recommendation, and obtain the recommendation result that reduces the proportion of useless users; including: introducing the latent representation learned from users and projects, adding it to the user representation and project representation respectively, and then dot-multiplying the two parts of information, outputting an n-dimensional vector, and then passing it to the prediction layer, which multiplies the weight with the n-dimensional vector, and then adds it with the user, project and global deviation to obtain the predicted rating score; The final result is the rating prediction scores of all items, generating the final item recommendation results.

2. The robust recommendation method for reducing the proportion of useless information under attack environment according to claim 1, characterized in that: The process of obtaining the feature representation of each user or item includes: Input the labeled useless information and real information into the prediction recommendation model to obtain all feature vectors of a user or item respectively; All feature vectors of a user or item are aggregated separately to obtain the feature representation of the user or item.

3. The robust recommendation method for reducing the proportion of useless information under attack environment according to claim 1, characterized in that: The process of predicting the sentiment polarity of the comment information is implemented using a double-layer LSTM model.

4. A robust recommendation system for reducing the proportion of useless information under attack, characterized in that: include: An acquisition module is configured to: acquire comment information in each entry, wherein each entry includes a rating score, comment information, an item ID, and a user ID; Extract the comment information of each entry, including: segmenting and embedding each comment information through bidirectional LSTM and Glove, storing and comparing using a dictionary, and generating positive or negative results for predicting the text; The sentiment analysis module is configured to: predict the sentiment polarity of the comment information to obtain the predicted sentiment polarity of each comment; compare the predicted sentiment polarity of each comment with the rating result corresponding to the comment, determine and mark useless information, including: in the scoring mechanism in the recommendation field, the rating score is a score in the interval [1, 5], 1 is the lowest score for a bad review, 5 is the highest score for a good review, and the middle value 3 is set in the interval of the rating score as the sentiment polarity dividing line, a rating greater than 3 is regarded as a positive sentiment, and a rating less than 3 is regarded as a negative sentiment; compare the rating sentiment polarity with the comment sentiment polarity, if the difference between the two sentiment polarities exceeds a set threshold, it is marked as useless information, otherwise, it is marked as true information; The prediction and recommendation module is configured as follows: based on the marked useless information and real information, a prediction and recommendation model is used to obtain the feature representation of each user or project; the attention mechanism is used to learn the weight of each comment, including: using the attention mechanism to learn to evaluate the weight of each comment to a certain extent, selecting comments that are meaningful to the project, and then aggregating these useful information to describe the project or user; in order to identify users who output meaningless comments, the unique information of each user and project, namely, ID information embedding, is introduced to simulate user quality, and the generalization effect of the model is improved by targeted information, and finally a discrete attention score is output, and the discrete attention score is normalized by using the softmax function; Combining the feature representation of each user or item respectively, including: multiplying the feature representation of each user or item by the weight of all comments respectively, and then obtaining the feature vector of the user or item respectively; Use the LFM component to complete the prediction rating and generate the final recommendation, and obtain the recommendation result that reduces the proportion of useless users; including: introducing the latent representation learned from users and projects, adding it to the user representation and project representation respectively, and then dot-multiplying the two parts of information, outputting an n-dimensional vector, and then passing it to the prediction layer, which multiplies the weight with the n-dimensional vector, and then adds it with the user, project and global deviation to obtain the predicted rating score; The final result is the rating prediction scores of all items, generating the final item recommendation results.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the robustness recommendation method for reducing the proportion of useless information under an attack environment as described in any one of claims 1 to 3 are implemented.

6. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the robustness recommendation method for reducing the proportion of useless information under an attack environment as described in any one of claims 1 to 3 are implemented.