CN112785331A - Injection attack resistant robust recommendation method and system combining evaluation text - Google Patents
Injection attack resistant robust recommendation method and system combining evaluation text Download PDFInfo
- Publication number
- CN112785331A CN112785331A CN202110018824.1A CN202110018824A CN112785331A CN 112785331 A CN112785331 A CN 112785331A CN 202110018824 A CN202110018824 A CN 202110018824A CN 112785331 A CN112785331 A CN 112785331A
- Authority
- CN
- China
- Prior art keywords
- evaluation
- model
- user
- personalized
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000011156 evaluation Methods 0.000 title claims abstract description 202
- 238000000034 method Methods 0.000 title claims abstract description 52
- 238000002347 injection Methods 0.000 title claims abstract description 35
- 239000007924 injection Substances 0.000 title claims abstract description 35
- 238000001514 detection method Methods 0.000 claims abstract description 43
- 238000012549 training Methods 0.000 claims abstract description 39
- 230000006870 function Effects 0.000 claims description 15
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 230000002159 abnormal effect Effects 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 5
- 230000015654 memory Effects 0.000 claims description 5
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 abstract description 17
- 238000004422 calculation algorithm Methods 0.000 abstract description 6
- 238000003062 neural network model Methods 0.000 abstract description 2
- 238000001914 filtration Methods 0.000 description 10
- 239000013598 vector Substances 0.000 description 6
- 230000006399 behavior Effects 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 230000003993 interaction Effects 0.000 description 4
- 230000001537 neural effect Effects 0.000 description 4
- 238000013459 approach Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 230000007547 defect Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000001737 promoting effect Effects 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 239000000243 solution Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 239000013604 expression vector Substances 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 230000000116 mitigating effect Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000001360 synchronised effect Effects 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/906—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Theoretical Computer Science (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Databases & Information Systems (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Business, Economics & Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- General Engineering & Computer Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Software Systems (AREA)
- Game Theory and Decision Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention designs a robust recommendation method and system for resisting injection attack by combining an evaluation text. The method comprises the following steps: 1) training an evaluation scoring prediction model by using a graph neural network model and using known information on a graph and calculated statistical information; 2) training a personalized evaluation text generation model by using a text generation model and a prediction score; 3) and training an attack detection model according to errors of the predicted value and the true value of the evaluation scoring prediction model and the personalized evaluation text generation model. The method integrates an evaluation scoring prediction model, a personalized evaluation text generation model and an attack detection model, so that three tasks are coordinated and mutually promoted. According to the scheme, evaluation text information is fully utilized, feedback data generated by a water army user is divided and utilized in a finer granularity mode, the influence of injection attack on recommendation accuracy in a commercial recommendation platform can be automatically relieved, and the robustness of a recommendation algorithm is improved.
Description
Technical Field
The method belongs to the technical field of information. Mainly aiming at a recommendation system in an e-commerce website, how to analyze user interest and commodity characteristics by combining evaluation text content, automatically identifying which evaluations come from injection attacks, and reducing the influence of the evaluations in the recommendation system. According to the method, three parts of personalized text content prediction, evaluation scoring prediction and attack detection are mutually promoted, so that the problem of attack injection in the e-commerce recommendation system can be automatically solved, and the robustness and accuracy of the recommendation system are improved.
Background
With the development of the e-commerce industry, more and more users choose to shop online. The E-commerce platform is provided with a large number of commodities which can be selected by a user, and in order to improve user experience and accelerate a purchasing decision process of the user, the E-commerce platform uses a recommendation system to guess user interest and provide personalized commodity recommendation for the user. It is very necessary for the e-commerce platform to ensure the accuracy of the recommendation system.
Of the recommended methods, collaborative filtering is a mainstream method, and can be further divided into memory-based collaborative filtering and model-based collaborative filtering. The collaborative filtering based on users and the collaborative filtering based on commodities are collectively called as collaborative filtering based on memories, and it is assumed that similar users have similar interests and similar commodities have similar characteristics, respectively. The model-based collaborative filtering is to express the scores given to the commodities by the users by using a matrix and decompose the scores into the product of two low-dimensional matrices, and essentially to perform matrix completion on the scores. Recently, the development of deep learning promotes the development of a recommendation method, and a series of deep collaborative filtering models improve the effect of a traditional recommendation method by jointly learning user vector representation representing user interest and commodity vector representation representing commodity characteristics. The users, the commodities and the feedback data of the users to the commodities in the E-commerce platform naturally form a graph structure, the graph neural network is an effective method for modeling the graph structure, and a series of deep collaborative filtering methods based on the graph neural network further improve the recommendation effect.
However, the user feedback data in e-commerce platforms is not all high quality and credible, and some merchants may hire the water army to issue false rated content, give unfair scores, for the purpose of promoting or defaulting targeted goods, a behavior called "injection attack". Some studies have shown that the presence of "injection attacks" can reduce the accuracy of the recommendation system, and thus impact market fairness.
With this in mind, work has been directed to mitigating the effects of "injection attacks" on the recommendation system. Some methods focus on how to improve recommendation system robustness by using robust models; some methods perform false detection on user feedback, remove false feedback, and then operate a recommendation system. The latest work (graphfi) indicates that the two methods respectively have the defects that the auxiliary information of the user is difficult to utilize and the normal data are judged as false data and deleted to possibly cause the model to be more and less trained, and the two methods are combined to respectively obtain the advantages and promote each other. Specifically, graphRfi is divided into two tasks of evaluation scoring prediction and water army detection, and is learned by Graph Convolutional Networks (Graph probabilistic Networks) and Neural Random Forests (Neural Random trees), which are jointly trained. In the evaluation scoring task, if the evaluation scoring of a user is obviously different from the predicted value, the user is probably a water army user; in a water army detection task, if a user is classified as a water army, the confidence of the user may be low, and the influence of the user in the evaluation scoring task needs to be reduced.
We have noted that the collaborative filtering approach described above in connection with graph neural networks is still inadequate in terms of the utilization of information on the graph. The method mainly utilizes node (user and commodity) attribute information and user scores on edges (feedback behaviors of the user on the commodity) to predict evaluation scores. However, the feedback of the user to the commodity not only scores the commodity, but also evaluates the commodity, and the commodity evaluation content contains rich text information. The rich text information can more accurately reflect the interests of the user and the characteristics of the commodities.
The method and the device have the key points that the existing graph-based recommendation algorithm is expanded, and the capturing capability of the recommendation algorithm on the evaluation text information is improved.
Disclosure of Invention
The invention provides an automatic robust recommendation method and system aiming at injection attack in a recommendation system and combining rich evaluation text content. The method integrates personalized text content prediction, evaluation scoring prediction and attack detection, and can promote each other.
In the following description, the data model of the present invention is denoted as (U ═ I, E), where U represents a user node set, I represents a commodity node set, and E represents an edge (relationship for users to generate feedback on commodities) set. The user node U belongs to U, the commodity node I belongs to I, the edge E belongs to E, and user feedback information (including evaluation text content C belongs to C and scoring R for commodities given by the user) and a specific time stamp T which generates feedback belong to T (T is a time stamp set) are arranged on the edge E.
The technical scheme adopted by the invention is as follows:
the invention firstly provides a robust recommendation model training method for resisting injection attack by combining an evaluation text, which comprises the following steps:
training an evaluation scoring prediction model;
training a personalized evaluation text generation model according to the evaluation score predicted by the evaluation score prediction model and the generated user representation and commodity representation;
training an attack detection model according to errors of predicted values and true values of the evaluation scoring prediction model and the personalized evaluation text generation model, and using the attack detection model for detection injection evaluation;
and according to the injection evaluation detection result output by the attack detection model, reducing the influence weight of corresponding evaluation in the evaluation scoring prediction model and the personalized evaluation text generation model.
Further, the training evaluation scoring prediction model is obtained by utilizing a graph neural network to train and obtain the representation of the user, the commodity and the feedback edge according to the user attribute, the commodity attribute, the evaluation content of the commodity given by the user, the scoring and the time for the user to generate feedback; the user attribute is a statistic that can reflect whether the user is abnormal.
Further, the above steps are divided into three tasks: the method comprises a personalized evaluation text generation task, an evaluation scoring prediction task and an attack detection task, wherein the three tasks are interactive and mutually promoted.
Specifically, the personalized evaluation text generation task can assist in evaluating the scoring prediction task; the evaluation scoring prediction task can guide a personalized evaluation text generation task; if the predicted comprehensive values of the first two tasks are obviously different from the actual values, the evaluation is probably injection evaluation; in the attack detection task, if one evaluation is classified as an injection evaluation, the influence of the evaluation in the first two tasks needs to be reduced (i.e., the influence weight is controlled).
Further, as the three tasks influence and promote each other, the invention designs a uniform loss function, so that the three tasks are uniform and carry out end-to-end cooperative training.
Specifically, for the evaluation scoring prediction task, we train the evaluation scoring prediction model S (U, I, C, R, T) → R 'so that the predicted score R' of the side to be tested approaches the true score
And when training the evaluation scoring prediction model, the influence weight of the false evaluation scoring needs to be reduced according to the output of the attack detection model.
Evaluation of the loss function of the scoring prediction model:
wherein p iseDenotes the probability, r ', of the edge e true'eA prediction score representing the edge e generated by the evaluation scoring prediction model,
represents an edge eThe true score of.
For the personalized evaluation text generation task, a personalized evaluation text generation model cg (U, I, C, R, T) → C 'is trained, so that the generated evaluation C' of the edge to be tested is close to the real evaluation
Further, a prediction score r 'generated by a rating and prediction model is added when the personalized rating text generation model is trained'eAnd (5) guiding. And when training the personalized evaluation text generation model, the influence weight of the false evaluation content needs to be reduced according to the output of the attack detection model. Meanwhile, an auxiliary model f for predicting, evaluating and scoring by evaluation content is introduced, individuation is carried out by using user representation and commodity representation generated by an evaluation and scoring prediction model, and the evaluation and scoring prediction model and an individualized evaluation text generation model are trained together by using the same set of user representation and commodity representation, so that an evaluation and scoring prediction task and an individualized evaluation text generation task are mutually cooperated.
Personalized evaluation of the loss function of the text generation model:
wherein p iseDenotes the probability, c ', of the edge e true'eAn evaluation representing the edge e generated by the personalized evaluation text generation model,
representing the true evaluation of the edge e.
For the attack detection task, updating the edge representation according to the difference between the content of the evaluation text predicted and generated on the edges of the former two tasks and the content of the real score and the real text (thereby influencing the training of the graph neural network in the evaluation score prediction task and the training of the generation model in the personalized evaluation text generation task in a new round of updating). Meanwhile, according to the updated edge representation, the attack detection model is trained to predict whether the edge comes fromAttack (p)eRepresenting the probability of the edge e being true).
Loss function (logarithmic loss) of attack detection model:
wherein, yeA label indicating whether an edge is real, | E | indicates the total number of edges, peRepresenting the probability of the edge e being true.
And the model overall loss function is formed by the loss functions of the three tasks together, and the three tasks are optimized uniformly in the training stage.
Specifically, the global loss function:
L＝Lrating+α·Lcontent_generate+β·Lfraudster
wherein, alpha and beta represent the over-parameters for controlling the three loss ratios.
On the basis of the training method, the invention provides an injection attack resistant robust recommendation method combined with an evaluation text.
Based on the same inventive concept, the invention also provides a robust recommendation system for resisting injection attack by combining with the evaluation text, which comprises:
the model training module is used for training a robust recommendation model which is combined with the evaluation text and is resistant to injection attack by adopting the method;
and the recommendation module is used for predicting the corresponding evaluation score of the user-commodity to be predicted by using the trained robust recommendation model which is combined with the evaluation text and resists the injection attack, and recommending the commodity according to the evaluation score.
The invention has the following beneficial effects:
the invention provides a unified training method integrating an evaluation scoring prediction task, a personalized evaluation text generation task and an attack detection task aiming at the problem that the accuracy of a recommendation system is reduced due to 'injection attack'. The existing methods only pay attention to improving the robustness of the recommendation algorithm, or the recommendation algorithm is applied after the attack is removed through a water army detection model, and the existing methods have the defects that the training of the model is worse due to the fact that the auxiliary information of a user is difficult to utilize, normal data are judged to be false and deleted. Recent work GraphRfi, although unified training is performed by combining an evaluation scoring prediction task and a water army detection task, has two problems that are not solved: first, some users, although hired by some merchants to issue "injection attack" ratings, may still have normal transaction, rating behavior, and all relevant feedback for users identified as "water force" is all discounted, losing data they normally feed back; second, in the utilization of feedback data, the model only utilizes the scoring data of the user on the goods, while in the e-market scenario, the text content of the user submitting an evaluation on the goods may reflect more information about the user's interest and about the characteristics of the goods, which was not utilized in the previous model.
The present technique focuses on improving the robustness of the recommendation algorithm. Compared with the prior art, the method fully utilizes the evaluation text information in the aspect of information source utilization; in the aspect of utilizing attack detection to reduce the weight, the method fully considers the characteristic that some 'water army' users still have normal feedback data on non-target commodities, and compared with the 'water army detection', the method uses the 'attack detection' to distinguish the feedback data generated by the 'water army' users in a finer granularity; in the aspect of model training, the problem of injection attack in the e-commerce recommendation system can be automatically relieved by mutually promoting three parts of evaluation scoring prediction, personalized evaluation text generation and attack detection, so that the robustness and accuracy of the recommendation system are improved.
Drawings
FIG. 1: and the evaluation scoring prediction model, the personalized evaluation text generation model and the attack detection model are used for generating an interaction frame diagram.
FIG. 2: e-business feedback network schematic diagram.
Detailed Description
The invention is further described in more detail below based on the e-commerce recommendation platform.
The invention relates to a robust recommendation method for evaluating text content in combination with injection attack, which comprises the following steps of:
1) according to a scoring matrix given to the commodity by the user and a time stamp for generating a score, a statistic amount which can reflect whether the user is abnormal or not is calculated as a user attribute.
2) According to the user attribute, the commodity attribute, the evaluation content and the score of the commodity given by the user and the time for the user to generate feedback, the representation of the user, the commodity and the feedback edge is obtained by utilizing the graph neural network training, and an evaluation score prediction model is obtained; and further predicting the corresponding score of the user-commodity pair to be tested by using the evaluation scoring prediction model (completing the scoring matrix).
3) And (2) utilizing a personalized evaluation text generation model to generate a personalized evaluation text (the generated evaluation content is related to the user interest and the commodity characteristics, specifically, the invention adds the user representation and the commodity representation in the evaluation scoring prediction model to the coding part of the text generation model to realize personalization), and adding the result of the evaluation scoring prediction in the step 2) to guide when training the generation model.
4) And (3) training an auxiliary model f for predicting the evaluation score by the evaluation content by using each piece of evaluation content c and the score r on the edge, so that f (c) is as close to r as possible. f (c) represents the evaluation score predicted from the evaluation content c. The auxiliary model f is used for optimizing user representation and commodity representation in the evaluation scoring prediction model through training of the personalized evaluation text generation model, so that the evaluation scoring prediction model is assisted to carry out evaluation scoring.
5) And comprehensively considering the difference between the generation result of the personalized evaluation text generation model and the evaluation scoring prediction model and the true value, performing an attack detection task, namely predicting whether the edge is false or not (the larger the difference is, the edge is more likely to be false, and the influence weight of the edge in the two previous models needs to be correspondingly reduced), generating a representation vector representing the difference, splicing the representation vector into the edge vector, acting on the personalized evaluation text generation model and the evaluation scoring prediction model, finally obtaining the corresponding score of the user-commodity pair to be predicted through the trained evaluation scoring prediction model, and recommending commodities according to the score.
The following provides a specific application example, which comprises the following steps:
1) figure 2 shows one example of an e-commerce feedback network. U in the figure1,u2,…,u8Representing a user node; i.e. i1,i2,…,i6A representative commodity node; connecting lines between the user nodes and the commodity nodes, wherein a solid line represents a real and known relationship of the user to the commodity feedback in the training set, and a dotted line represents the preference degree of the user to be predicted in the testing set to the commodity; the edge is provided with a timestamp for the occurrence of the feedback relationship; on the edge of cu,i，ru,iRespectively representing the evaluation text published by the user on the commodity and the score given by the user on the commodity.
2) And calculating some statistical information for the user according to the feedback score and the occurrence time of the feedback behavior to serve as attributes reflecting whether the user is abnormal or not. Such as the total number of items rated by the user, the proportion of scores from 1 to 5 scored by the user, the entropy of scores scored by the user, the proportion of "praise" to all ratings of the user, the interval of time between feedback actions by the user, and so forth.
3) And training an evaluation scoring prediction model by using the attributes of the user and the commodity nodes on the graph, the feedback relation and the occurrence time of the user to the commodity, the specific scoring of the user to the commodity and the rich information of the evaluation text content fed back by the user. The invention takes an SSG model (Set-Sequence-Graph: A Multi-View Approach expanding Reviews for recommendation. in CIKM.2020.) as an example to complete the evaluation scoring matrix. The SSG is the latest work of comprehensively utilizing the attributes of the user and the commodity nodes, the feedback relationship and the occurrence time of the feedback behavior, and publishing the evaluation text content by the user to carry out the commodity scoring prediction by the user. It utilizes multi-view learning to evaluate published by users from three views: viewing is carried out by integrating the viewing angle, the sequence viewing angle and the drawing viewing angle, so that the expression vectors of users, commodities and feedback are learned, and the scoring of a target user-commodity pair is further predicted. The set view angle is to regard different evaluations published by the same user (or different evaluations under the same commodity) as a set, and then calculate the unified representation of the set as the set representation of the user (or the commodity) through an attention network; the sequence view angle is that the evaluations published by the same user (or different evaluations under the same commodity) are arranged into a sequence according to the time sequence, and then the unified representation of the sequence is calculated through an LSTM (long short term memory network) to be used as the sequence representation of the user (or the commodity); in the view angle of the graph, the network relation between users and commodities is utilized, the nodes and the side representation are subjected to sufficient information interaction through an RGAT (graph attention network introducing evaluation), and the representation vectors of the users, the commodity nodes and the side evaluation on the graph are comprehensively learned to be represented under the view angle of the graph. The representations of the three views are combined in a certain proportion by hyper-parameters to make a prediction of the end user's score on the commodity.
4) And carrying out personalized evaluation generation by using the personalized evaluation text generation model, and simultaneously keeping synchronous optimization with the evaluation scoring prediction task. The invention can fully utilize the evaluation scoring prediction and the evaluation text content to generate the connection between the two tasks, so that the two tasks are mutually promoted. Furthermore, the invention can also utilize a Transformer architecture to introduce user interest representation in the Transformer so as to generate more accurate and more personalized evaluation text content.
5) Attack detection task, and interaction with the first two tasks (evaluation scoring prediction task and personalized evaluation text generation task), we can adopt a method similar to GraphRfi (GCN-Based User retrieval Learning for unity road repair communication and Fraudster detection. in KDD.2019.) and expand. Specifically, the graphRfi model is divided into two tasks of evaluation scoring prediction and water army detection, and is learned by Graph Convolutional Networks (Graph probabilistic Networks) and Neural Random Forests (Neural Random trees), which are jointly trained. In the evaluation scoring task, if the evaluation scoring of a user is obviously different from the predicted value, the user is probably a water army user; in a water army detection task, if a user is classified as a water army, the confidence of the user may be low, and the influence of the user in the evaluation scoring task needs to be reduced. As described above, GraphRfi is insufficient in utilizing user feedback information, and does not utilize specific text content of the user's evaluation of the merchandise release. According to the method, evaluation text information is further introduced on the basis of GraphRfi, a personalized evaluation text generation task is added, and the evaluation text generation task and other two tasks are trained jointly and mutually promoted. In this component, we view the first two tasks (the evaluation scoring prediction task and the personalized evaluation text generation task) as an overall scoring module, analogizing to the evaluation scoring module in graphfi, as shown in fig. 1. And the scoring module further interacts with the attack detection task.
Based on the same inventive concept, another embodiment of the present invention provides a robust recommendation system against injection attack in combination with rating text, comprising:
the model training module is used for training a robust recommendation model which is used for resisting injection attack and combined with the evaluation text by adopting the method;
and the recommendation module is used for predicting the corresponding evaluation score of the user-commodity to be predicted by using the trained robust recommendation model which is combined with the evaluation text and resists the injection attack, and recommending the commodity according to the evaluation score.
Based on the same inventive concept, another embodiment of the present invention provides an electronic device (computer, server, smartphone, etc.) comprising a memory storing a computer program configured to be executed by the processor, and a processor, the computer program comprising instructions for performing the steps of the inventive method.
Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium (e.g., ROM/RAM, magnetic disk, optical disk) storing a computer program, which when executed by a computer, performs the steps of the inventive method.
The present invention is not limited to the manner described in the above embodiments, such as:
1. the statistic for reflecting whether the user is abnormal or not can be limited to the total number of commodities evaluated by the user, the ratio of scores from 1 to 5 given by the user, the entropy value given by the user, the ratio of 'praise' of all evaluations of the user, the interval time of feedback actions of the user and the like, and also can comprise any other statistic which is calculated according to the node attribute, the edge structure and the edge timestamp on the graph and can possibly reflect the characteristics of the water army of the user;
2. the model for making the evaluation scoring prediction from the existing on-graph information may use other graph neural networks or non-graph neural network models, including other graph representation learning models that partially utilize on-graph information (which is not fully utilized on existing on-graph information).
3. Other text generation models such as LSTM, Transformer and expansion thereof can be adopted, and other modes except multitask learning and bidirectional learning can be adopted to carry out mutual promotion and common optimization on the personalized text generation task and the evaluation scoring task.
4. Besides the method for the invention, other three task interaction modes can be adopted to carry out the common optimization of the three tasks.
The above embodiments are only intended to illustrate the technical solution of the present invention and not to limit the same, and a person skilled in the art can modify the technical solution of the present invention or substitute the same without departing from the principle and scope of the present invention, and the scope of the present invention should be determined by the claims.
Claims (10)
1. A robust recommendation model training method for resisting injection attack combined with an evaluation text is characterized by comprising the following steps:
training an evaluation scoring prediction model;
training a personalized evaluation text generation model according to the evaluation score predicted by the evaluation score prediction model and the generated user representation and commodity representation;
training an attack detection model according to errors of predicted values and true values of the evaluation scoring prediction model and the personalized evaluation text generation model, and using the attack detection model for detection injection evaluation;
and according to the injection evaluation detection result output by the attack detection model, reducing the influence weight of corresponding evaluation in the evaluation scoring prediction model and the personalized evaluation text generation model.
2. The method according to claim 1, wherein the training of the evaluation scoring prediction model is based on user attributes, commodity attributes, evaluation contents of commodities given by users, scoring and time for users to generate feedback, and the evaluation scoring prediction model is obtained by training with a graph neural network to obtain representations of users, commodities and feedback edges; the user attribute is a statistic that can reflect whether the user is abnormal.
3. The method according to claim 1, characterized in that it comprises three tasks: an evaluation scoring prediction task realized by the evaluation scoring prediction model, a personalized evaluation text generation task realized by the personalized evaluation text generation model, and an attack detection task realized by the attack detection model; the three tasks interact, promote each other: generating a task auxiliary evaluation scoring prediction task by using the personalized evaluation text; the evaluation scoring prediction task guides a personalized evaluation text generation task; in the attack detection task, if one evaluation is classified as an injection evaluation, the influence of the evaluation in the first two tasks is reduced.
4. The method of claim 3, wherein the three tasks are uniformly co-trained end-to-end using a uniform loss function, wherein the overall loss function is:
L＝Lrating+α·Lcontent_generate+β·Lfraudster
wherein L isratingTo evaluate the loss function of the scoring prediction model, Lcontent_generateGenerating a loss function of the model for the personalized evaluation of the text, LfraudsterAlpha and beta represent the hyperparameters controlling the three loss proportions for the loss function of the attack detection model.
5. The method of claim 4, wherein the loss function of the evaluation scoring prediction model is:
wherein p iseDenotes the probability, r ', of the edge e true'eA prediction score representing the edge e generated by the evaluation scoring prediction model,
representing the true score of the edge e.
6. The method according to claim 4, characterized in that, when training the personalized evaluation text generation model, an auxiliary model f for predicting evaluation scores by evaluation contents is introduced, and personalization is performed by using user representation and commodity representation generated by the evaluation score prediction model, and the evaluation score prediction model and the personalized evaluation text generation model are trained together by using the same set of user representation and commodity representation, so that an evaluation score prediction task and a personalized evaluation text generation task are coordinated with each other; the personalized evaluation text generation model comprises the following steps:
cg(U，I，C，R，T)→C′
the system comprises a user node set, a commodity node set, an evaluation text content set, a timestamp set and an evaluation text content set, wherein U represents the user node set, I represents the commodity node set, C represents the evaluation text content set, R represents the scoring set of commodities given by a user, T is the timestamp set, and C' is the generated evaluation of a to-be-tested edge;
the loss function of the personalized evaluation text generation model is as follows:
wherein p iseDenotes the probability, c ', of the edge e true'eAn evaluation representing the edge e generated by the personalized evaluation text generation model,
representing the true evaluation of the edge e.
7. The method of claim 4, wherein the loss function of the attack detection model is:
wherein, yeA label indicating whether an edge is real, | E | indicates the total number of edges, peRepresenting the probability of the edge e being true.
8. A robust recommendation method for resisting injection attack in combination with an evaluation text is characterized by comprising the following steps:
and predicting the corresponding evaluation score of the user-commodity to be predicted by using the injection attack resistant robust recommendation model combined with the evaluation text after the training is finished by the method of any one of claims 1 to 7, and recommending the commodity according to the evaluation score.
9. A robust recommendation system for injection attack resistance in conjunction with evaluation text, comprising:
the model training module is used for training a robust recommendation model which is used for resisting injection attack and combined with the evaluation text by adopting the method of any one of claims 1-7;
and the recommendation module is used for predicting the corresponding evaluation score of the user-commodity to be predicted by using the trained robust recommendation model which is combined with the evaluation text and resists the injection attack, and recommending the commodity according to the evaluation score.
10. An electronic apparatus, comprising a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program comprising instructions for performing the method of any of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110018824.1A CN112785331A (en) | 2021-01-07 | 2021-01-07 | Injection attack resistant robust recommendation method and system combining evaluation text |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110018824.1A CN112785331A (en) | 2021-01-07 | 2021-01-07 | Injection attack resistant robust recommendation method and system combining evaluation text |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112785331A true CN112785331A (en) | 2021-05-11 |
Family
ID=75756038
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110018824.1A Pending CN112785331A (en) | 2021-01-07 | 2021-01-07 | Injection attack resistant robust recommendation method and system combining evaluation text |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112785331A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117473284A (en) * | 2023-11-20 | 2024-01-30 | 灏冉舟网络有限公司 | Three-party transaction platform based on artificial intelligence and method thereof |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130204833A1 (en) * | 2012-02-02 | 2013-08-08 | Bo PANG | Personalized recommendation of user comments |
| US20130325573A1 (en) * | 2012-06-04 | 2013-12-05 | Neowiz Games Corporation | Method, apparatus, and recording medium implementing mobile application marketing |
| CN105574003A (en) * | 2014-10-10 | 2016-05-11 | 华东师范大学 | Comment text and score analysis-based information recommendation method |
| JP2017027480A (en) * | 2015-07-24 | 2017-02-02 | 大学共同利用機関法人情報・システム研究機構 | Item recommendation system and item recommendation method |
| JP2018063484A (en) * | 2016-10-11 | 2018-04-19 | 凸版印刷株式会社 | User's evaluation prediction system, user's evaluation prediction method and program |
| CN108470052A (en) * | 2018-03-12 | 2018-08-31 | 南京邮电大学 | A kind of anti-support attack proposed algorithm based on matrix completion |
| JP2019114053A (en) * | 2017-12-22 | 2019-07-11 | Kddi株式会社 | Recommending apparatus, recommending method and recommending program |
| CN110210933A (en) * | 2019-05-21 | 2019-09-06 | 清华大学深圳研究生院 | A kind of enigmatic language justice recommended method based on generation confrontation network |
| CN111222332A (en) * | 2020-01-06 | 2020-06-02 | 华南理工大学 | Commodity recommendation method combining attention network and user emotion |
| CN111275521A (en) * | 2020-01-16 | 2020-06-12 | 华南理工大学 | Commodity recommendation method based on user comment and satisfaction level embedding |
| CN111414549A (en) * | 2019-05-14 | 2020-07-14 | 北京大学 | Intelligent general assessment method and system for vulnerability of recommendation system |
| CN111639184A (en) * | 2020-06-01 | 2020-09-08 | 复旦大学 | Detection system for tendency inconsistency of scores and comment contents |
| CN111930926A (en) * | 2020-08-05 | 2020-11-13 | 南宁师范大学 | Personalized recommendation algorithm combined with comment text mining |
| US10868888B1 (en) * | 2015-12-02 | 2020-12-15 | Zeta Global Corp. | Method and apparatus for real-time personalization |
-
2021
- 2021-01-07 CN CN202110018824.1A patent/CN112785331A/en active Pending
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130204833A1 (en) * | 2012-02-02 | 2013-08-08 | Bo PANG | Personalized recommendation of user comments |
| US20130325573A1 (en) * | 2012-06-04 | 2013-12-05 | Neowiz Games Corporation | Method, apparatus, and recording medium implementing mobile application marketing |
| CN105574003A (en) * | 2014-10-10 | 2016-05-11 | 华东师范大学 | Comment text and score analysis-based information recommendation method |
| JP2017027480A (en) * | 2015-07-24 | 2017-02-02 | 大学共同利用機関法人情報・システム研究機構 | Item recommendation system and item recommendation method |
| US10868888B1 (en) * | 2015-12-02 | 2020-12-15 | Zeta Global Corp. | Method and apparatus for real-time personalization |
| JP2018063484A (en) * | 2016-10-11 | 2018-04-19 | 凸版印刷株式会社 | User's evaluation prediction system, user's evaluation prediction method and program |
| JP2019114053A (en) * | 2017-12-22 | 2019-07-11 | Kddi株式会社 | Recommending apparatus, recommending method and recommending program |
| CN108470052A (en) * | 2018-03-12 | 2018-08-31 | 南京邮电大学 | A kind of anti-support attack proposed algorithm based on matrix completion |
| CN111414549A (en) * | 2019-05-14 | 2020-07-14 | 北京大学 | Intelligent general assessment method and system for vulnerability of recommendation system |
| CN110210933A (en) * | 2019-05-21 | 2019-09-06 | 清华大学深圳研究生院 | A kind of enigmatic language justice recommended method based on generation confrontation network |
| CN111222332A (en) * | 2020-01-06 | 2020-06-02 | 华南理工大学 | Commodity recommendation method combining attention network and user emotion |
| CN111275521A (en) * | 2020-01-16 | 2020-06-12 | 华南理工大学 | Commodity recommendation method based on user comment and satisfaction level embedding |
| CN111639184A (en) * | 2020-06-01 | 2020-09-08 | 复旦大学 | Detection system for tendency inconsistency of scores and comment contents |
| CN111930926A (en) * | 2020-08-05 | 2020-11-13 | 南宁师范大学 | Personalized recommendation algorithm combined with comment text mining |
Non-Patent Citations (7)
| Title |
|---|
| RUI LIU等: "STARS: Defending against Sockpuppet-Based Targeted Attacks on Reviewing Systems", 《ACM TRANSACTIONS ON INTELLIGENT SYSTEMS AND TECHNOLOGY》, vol. 11, no. 5, 24 July 2020 (2020-07-24), pages 56 * |
| SHIJIE ZHANG等: "GCN‑Based User Representation Learning for Unifying Robust Recommendation and Fraudster Detection", 《PROCEEDINGS OF THE 43RD INTERNATIONAL ACM SIGIR CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION RETRIEVAL》, 25 July 2020 (2020-07-25), pages 689 - 698, XP058748006, DOI: 10.1145/3397271.3401165 * |
| SONG JUNSHUAI等: "《PoisonRec: An Adaptive Data Poisoning Framework for Attacking Black-box Recommender Systems》", 《2020 IEEE 36TH INTERNATIONAL CONFERENCE ON DATA ENGINEERING》, 27 May 2020 (2020-05-27), pages 157 - 168 * |
| VENUGOPALAN, M 等: "Rating Prediction Model for Reviews Using a Novel Weighted Textual Feature Method", 《RECENT FINDINGS IN INTELLIGENT COMPUTING TECHNIQUES, 》, vol. 709, 4 November 2018 (2018-11-04), pages 177 - 190 * |
| 刘鑫等: "限制项目关注度的鲁棒推荐算法", 《天津理工大学学报》, vol. 36, no. 01, 15 February 2020 (2020-02-15), pages 1 - 5 * |
| 司明丹: "推荐系统中虚假评价用户识别方法研究", 《中国优秀硕士学位论文全文数据库 (信息科技辑)》, no. 2, 15 February 2020 (2020-02-15), pages 138 - 2282 * |
| 孙俊等: "基于深度神经网络的带有评论文本的推荐模型", 《信息与电脑(理论版)》, vol. 32, no. 01, 10 January 2020 (2020-01-10), pages 123 - 125 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117473284A (en) * | 2023-11-20 | 2024-01-30 | 灏冉舟网络有限公司 | Three-party transaction platform based on artificial intelligence and method thereof |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| KR102233589B1 (en) | Modeling method and device for evaluation model | |
| Nagengast et al. | The great collapse in value added trade | |
| CN108665339B (en) | E-commerce product reliability index based on subjective emotion measure and implementation method thereof | |
| CN110263244A (en) | Content recommendation method, device, storage medium and computer equipment | |
| KR20180041174A (en) | Risk Assessment Methods and Systems | |
| Wu et al. | Deep censored learning of the winning price in the real time bidding | |
| CN109800976A (en) | Investment decision methods, device, computer equipment and storage medium | |
| CN110717654B (en) | Product quality evaluation method and system based on user comments | |
| CN105096152A (en) | Commodity popularity-based operation execution method and device | |
| CN113537807B (en) | Intelligent wind control method and equipment for enterprises | |
| CN109583782B (en) | Automobile financial wind control method supporting multiple data sources | |
| EP2939196A1 (en) | Techniques for measuring video profit | |
| CN112116387A (en) | Information prediction method based on causal relationship | |
| Xu et al. | [Retracted] Research on the Construction of Crossborder e‐Commerce Logistics Service System Based on Machine Learning Algorithms | |
| CN112785331A (en) | Injection attack resistant robust recommendation method and system combining evaluation text | |
| Yuping et al. | New methods of customer segmentation and individual credit evaluation based on machine learning | |
| CN110738565A (en) | Real estate finance artificial intelligence composite wind control model based on data set | |
| Lo et al. | A fuzzy group-preferences analysis method for new-product development | |
| Gabbriellini et al. | A micro study on the evolution of arguments in amazon. com’s reviews | |
| CN106022915A (en) | Enterprise credit risk assessment method and apparatus | |
| CN109359865A (en) | Product evaluation method, apparatus, computer equipment and storage medium | |
| CN114926208A (en) | User demand data analysis method and system for product improvement strategy formulation | |
| Bennani et al. | Does central bank communication signal future monetary policy? The case of the ECB | |
| CN115860889A (en) | Financial loan big data management method and system based on artificial intelligence | |
| CN107645412B (en) | Web service combination multi-target verification method in open environment |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |