CN114254187B - Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training - Google Patents
Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training Download PDFInfo
- Publication number
- CN114254187B CN114254187B CN202111506522.5A CN202111506522A CN114254187B CN 114254187 B CN114254187 B CN 114254187B CN 202111506522 A CN202111506522 A CN 202111506522A CN 114254187 B CN114254187 B CN 114254187B
- Authority
- CN
- China
- Prior art keywords
- user
- loss function
- article
- attribute data
- diagram
- 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.)
- Active
Links
- 238000012549 training Methods 0.000 title claims abstract description 75
- 238000000034 method Methods 0.000 title claims abstract description 44
- 230000009467 reduction Effects 0.000 title claims abstract description 37
- 230000003993 interaction Effects 0.000 claims abstract description 81
- 230000006870 function Effects 0.000 claims description 151
- 238000010586 diagram Methods 0.000 claims description 67
- 239000013598 vector Substances 0.000 claims description 41
- 238000012545 processing Methods 0.000 claims description 29
- 230000008569 process Effects 0.000 claims description 17
- 230000003044 adaptive effect Effects 0.000 claims description 14
- 239000011159 matrix material Substances 0.000 claims description 14
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 238000004220 aggregation Methods 0.000 claims description 12
- 238000004891 communication Methods 0.000 claims description 12
- 238000005457 optimization Methods 0.000 claims description 12
- 230000002776 aggregation Effects 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000012163 sequencing technique Methods 0.000 claims description 7
- 230000004931 aggregating effect Effects 0.000 claims description 6
- 210000005036 nerve Anatomy 0.000 claims description 5
- 238000003860 storage Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 abstract description 4
- 238000005520 cutting process Methods 0.000 abstract description 3
- 230000002028 premature Effects 0.000 abstract description 3
- 238000013527 convolutional neural network Methods 0.000 description 7
- 230000006399 behavior Effects 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 230000009977 dual effect Effects 0.000 description 4
- 238000001914 filtration Methods 0.000 description 3
- 238000003058 natural language processing Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 230000001537 neural effect Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- 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
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Evolutionary Biology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Filters That Use Time-Delay Elements (AREA)
- Noise Elimination (AREA)
- Soundproofing, Sound Blocking, And Sound Damping (AREA)
Abstract
The invention discloses a recommendation method, a system, electronic equipment and a medium based on self-adaptive noise reduction training, which relate to the technical field of computers, the loss function is truncated or re-weighted during the training phase, and the noise reduction is automatically performed during the training phase of the model. Compared with the prior art, the prior recommendation model does not consider that the noise problem accompanied by implicit feedback is processed in the training stage; the invention carries out the cutting or weighting treatment on the loss function so as to optimize the scoring function, thereby greatly reducing the noise influence of the false positive interaction on the training model, reducing the workload of manually screening out the false positive interaction, improving the recommended accuracy and relieving the premature fitting phenomenon of the training model.
Description
Technical Field
The invention relates to the technical field of computers, in particular to a recommendation method, a recommendation system, electronic equipment and a recommendation medium based on self-adaptive noise reduction training.
Background
The recommendation system is widely applied in real life and is also an important practical application of natural language processing (natural language processing, NLP). Thanks to the rapid development of the recommendation system, each large network platform has obtained revolutionary progress in real scenes such as search engines, e-commerce shopping, social media, news portals and the like, and almost every service providing recommendation functions for users is provided with a corresponding recommendation system.
When the recommendation model is trained, a large amount of implicit feedback is often adopted. While implicit feedback alleviates the data sparsity problem, it contains many false positive interactions: i.e. the user interacts with the item or article, the final feedback of the user is that the item is disliked. In other words, implicit feedback is substantially accompanied by noise when reflecting the actual satisfaction of the user. For example, in online shopping, most click-through does not translate into purchasing behavior, even though many purchasing behaviors are accompanied by negative ratings, but the model takes all implicit feedback into account during the training phase, i.e., all click-through behavior. In current recommendation models, there are few considerations of how to deal with the noise problems created by using implicit feedback during the training phase. However, theoretical studies have found that: during the training phase of the model, particularly during the early phase, false positive interactions can produce large loss values; in order to reduce the occurrence of the over-fitting phenomenon, the model should learn more about the true positive interaction features during the early training phase of the model.
Therefore, how to filter out noise generated by using implicit feedback in the training phase, that is, to automatically reduce noise for false positive interaction in the training phase, is considered to be a solution idea for improving the recommendation effect.
Disclosure of Invention
The invention aims to provide a recommendation method, a system, electronic equipment and a medium based on self-adaptive noise reduction training, which are used for solving the problem of how to reduce noise influence encountered by a recommendation model in the training process, so that the recommendation accuracy is improved.
The technical aim of the invention is realized by the following technical scheme:
in a first aspect, a recommendation method based on adaptive noise reduction training is provided, including the following steps:
Acquiring an interaction data set of a user and an article, constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article in the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram;
aggregating the characteristics of nodes adjacent to each entity node in the knowledge graph by using a graph convolution neural network, and performing splicing processing on the characteristics of the entity node and the characteristics of the nodes adjacent to the entity node through a full-connection layer to obtain the characteristic vector of the user and the object;
in the training stage of the nerve model, learning a scoring function through a loss function, and carrying out truncation or re-weighting treatment on the loss function so as to optimize the scoring function;
And carrying out inner product calculation on the feature vectors of the user and the article by using the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommendation result.
In the graph convolution neural network, the invention carries out truncation processing or re-weighting processing on the loss function in the training stage, and automatically carries out optimization processing on the scoring function in the training stage of the model. Compared with the prior art, the prior recommendation model does not consider that the noise problem accompanied by implicit feedback is processed in the training stage; the invention carries out the cutting or weighting treatment on the loss function to optimize the scoring function, thereby greatly reducing the noise influence of the false positive interaction on the training model, reducing the workload of manually screening the false positive interaction, improving the recommendation accuracy and relieving the premature fitting phenomenon of the training model.
Further, an entity alignment matrix of the user and the article is constructed according to the data interaction diagram and the attribute data relation diagram, and the data interaction diagram of the user and the article and the attribute data relation diagram of the user and the article are utilized to form a knowledge diagram comprising the user, the article and the attribute data.
Further, before aggregation is performed by using the graph convolution neural network, the embedded layer of the graph convolution neural network is used for performing dimension reduction processing on the high-dimensional vector in the knowledge graph to obtain a low-dimensional vector.
Further, a dual permutation loss function is employed to optimize vectors in the embedded layer.
Further, the implementation of the truncation process for the loss function is as follows:
And learning the scoring function by using the loss function, setting a truncated value of the loss function, setting the result calculated by the loss function to be zero when the loss value calculated by the loss function is larger than the truncated value and real interaction occurs between the user and the article, otherwise, keeping the result calculated by the loss function.
Further, the implementation of the re-weighting of the loss function is as follows:
the scoring function is learned using the loss function, and weights are added to the loss function to reduce the loss value of the loss function.
Further, the loss function is subjected to iterative training, so that the loss value calculated by the loss function is minimized, and meanwhile, the loss function is subjected to noise reduction in the iterative training process, and an optimized scoring function is obtained.
In a second aspect, there is provided a recommendation system based on adaptive noise reduction training, comprising,
The generating unit is used for acquiring an interaction data set of the user and the article and constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article into the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram;
The aggregation unit is used for aggregating the characteristics of the nodes adjacent to each entity node in the knowledge graph by using the graph convolution neural network, and performing splicing processing on the characteristics of the entity nodes and the characteristics of the nodes adjacent to the entity nodes through the full connection layer to obtain the characteristic vectors of the user and the article;
The optimizing unit is used for learning the scoring function through the loss function in the training stage of the nerve model, and carrying out truncation or re-weighting treatment on the loss function so as to optimize the scoring function;
And the recommending unit is used for carrying out inner product calculation on the feature vectors of the user and the article by utilizing the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommending result.
In a third aspect, an electronic device is provided, comprising: the device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus; the memory is used for storing a computer program; the processor is configured to execute the program stored in the memory, and implement the recommendation method based on adaptive noise reduction training according to the first aspect.
In a fourth aspect, a computer readable storage medium is provided, storing a computer program, which when executed by a processor implements the recommendation method based on adaptive noise reduction training according to the first aspect.
Compared with the prior art, the invention has the following beneficial effects:
According to the invention, on a recommendation system based on a graph convolution neural network, a truncation process or a re-weighting process is carried out on a loss function in a training stage, and a noise reduction process is automatically carried out in a training stage of a model. Compared with the prior art, the prior recommendation model does not consider that the noise problem accompanied by implicit feedback is processed in the training stage; by cutting off or weighting the loss function, the invention can greatly reduce the noise influence of the false positive interaction line on the training model, not only can reduce the workload of manually screening out the false positive interaction line, but also can improve the recommendation accuracy and relieve the too-fitting phenomenon of the training model.
Drawings
The accompanying drawings, which are included to provide a further understanding of embodiments of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:
FIG. 1 is a flowchart of a recommendation method according to an embodiment of the present invention;
FIG. 2 is a block diagram of a recommendation system according to an embodiment of the present invention;
fig. 3 is a block diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
For the purpose of making apparent the objects, technical solutions and advantages of the present invention, the present invention will be further described in detail with reference to the following examples and the accompanying drawings, wherein the exemplary embodiments of the present invention and the descriptions thereof are for illustrating the present invention only and are not to be construed as limiting the present invention.
It will be understood that when an element is referred to as being "mounted" or "disposed" on another element, it can be directly on the other element or be indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly or indirectly connected to the other element.
It is to be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like are merely for convenience in describing and simplifying the description based on the orientation or positional relationship shown in the drawings, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus are not to be construed as limiting the invention.
Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more such feature. In the description of the present invention, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.
Embodiment one:
as shown in fig. 1, a first embodiment of the present application provides a recommendation method based on adaptive noise reduction training, which includes the following steps:
S1, acquiring an interaction data set of a user and an article, constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article in the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram;
s2, aggregating the characteristics of nodes adjacent to each entity node in the knowledge graph by using a graph convolution neural network, and performing splicing processing on the characteristics of the entity node and the characteristics of the nodes adjacent to the entity node through a full connection layer to obtain the characteristic vectors of the user and the object;
S3, learning a scoring function through a loss function in a neural model training stage, and performing truncation or re-weighting treatment on the loss function so as to optimize the scoring function;
And S4, performing inner product calculation on the feature vectors of the user and the article by using the scoring function after optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommendation result.
Specifically, in step S1, a user and article interaction data set is collected, and a data interaction graph G1 of the user and the article is constructed according to whether the user and the article interact, specifically: the method comprises the steps of collecting and processing public data, extracting interaction relation between a user and an article, and constructing a user-article interaction graph G1 which is expressed as { (U, y ui, I) |u epsilon U, I epsilon I, wherein U and I respectively represent a user data set and an article data set, and y ui＝1,yui =0 respectively represent interaction or no interaction between the user U and the article I.
Adding attribute data of the user and the object to form an attribute data relation graph G2 of the user-attribute and the object-attribute, wherein the purpose is to further aggregate attribute information, and the specific implementation of the attribute data relation graph G2 is as follows: the new attribute data relationship graph G2 is formed by the attribute data of the user (for example, additional information such as age, international, occupation, etc.), the attribute data of the object (for example, additional information such as subject, actor, director, etc.), and the user and the object. It is expressed as: { (h, R, t) |h, t ε, R ε R } where h, t are head and tail entities, respectively, R are relationship entities, ε and R are entity sets and relationship sets, respectively, such as: (Hugh Jackman, actor Of, logan) means "Hugh Jackman is an Actor Of the movie Logan".
Using the user and entity alignment matrix, the article and the entity alignment matrix, forming a new knowledge graph G by G1 and G2, wherein the user, the article and the attribute are all expressed in the knowledge graph in the form of entity nodes, and the method is specifically realized as follows: and establishing an article and entity alignment matrix A= { (I, e) |e epsilon, I epsilon I } and aligning the article I in the user and article data interaction diagram G1 with the entity e in the article attribute knowledge diagram G2 one by one. The object node I exists in the G1 and the G2 at the same time, I in the G2 is not an object but an entity, and the entity set U in the G2 is far greater than the object set I in the data interaction graph G1 of the user and the object, so that the I is used in the two graphs at the same time and the object-entity alignment matrix is used for conversion;
The attribute data knowledge graph G2 of the article and the data interaction graph G1 of the user and the article can be seamlessly combined into a new knowledge graph G by utilizing the article and entity alignment matrix A, and the new knowledge graph G, G and G2 are similar in representation form and are expressed as { (h, R, t) |h, t epsilon ', R epsilon R' }, epsilon '=epsilon U and R' =R U { inter }.
In step S2, the neighboring node features of each entity node are aggregated on the knowledge graph G by using the graph convolutional neural network specifically as follows: computing an aggregate representation of neighbor vectors for an entity e: Where e t is the vector representation of the tail entity corresponding to the vector representation e h using e as the head entity, and N h = { (h, r, t) | (h, r, t) ∈g } represents the triplet set using e as the head entity.
The aggregated features pass through a full connection layer, and the original features of the entity nodes (namely, the self features before information aggregation) and the features of the nodes adjacent to the entity nodes are combined together while the self features of the entity nodes are reserved, wherein the method specifically comprises the following steps: aggregating the information of the entity itself and the surrounding node information together, utilizing a functionAggregation of features is achieved, and ReLU () is an activation function.
In step S3, the loss function subjected to the noise reduction processing is truncated or re-weighted, so that the loss function can minimize the noise generated by the implicit feedback to the maximum and automatically.
In step S4, first, the feature vector representations of user u and item i are connected using a layer aggregation mechanism: wherein L represents that the vector of the user u and the object i aggregates surrounding L-level node information; the two multi-layer vector representations are then inner-product to predict the matching score of user u and item i:
Will be And (5) sorting, and recommending a plurality of articles with highest matching scores to the user according to the sorting result.
The recommendation method of the application can be applied to movie recommendation, book recommendation and music recommendation or platform shopping recommendation, such as that a user only clicks a movie and does not watch at all, or finishes the movie and gives negative evaluation, which is considered to be false positive interaction and has no training meaning, and filtering should be performed in the training stage. The noise reduction treatment is carried out in the model training stage, and the effect of the false positive interaction behavior on the model is actually reduced to the minimum to the maximum extent and automatically. Thus, not only is manpower saved to screen the data set, but also the following steps are found: the recommended effect after the noise reduction treatment is better improved compared with a recommended model without the noise reduction treatment.
In a further embodiment of the present application, an entity alignment matrix of the user and the item is constructed according to the data interaction diagram and the attribute data relationship diagram, and the data interaction diagram of the user and the item and the attribute data relationship diagram of the user and the item are used to form a knowledge diagram including the user, the item and the attribute data.
In a further embodiment of the present application, the high-dimensional vectors in the knowledge graph are reduced to low-dimensional vectors by using an embedding layer of the graph convolutional neural network before aggregation by using the graph convolutional neural network.
Specifically, the high-dimensional low-density entity vector and the low-dimensional vector of the relation vector in the knowledge graph G are reduced in dimension so as to facilitate feature aggregation and correlation comparison, and the low-dimensional vector specifically comprises:
the triples (h, r, t) exist in the knowledge graph G, wherein the embedments of h, t and r are e h,et∈Rd,er∈Rk respectively, and the vector dimensions are not consistent and d and k respectively;
Therefore, while preserving the graph structure of G, parameterizing entities and relationships into a low-dimensional vector representation requires building entity and relationship embeddings in entity space and relationship space, respectively: in this case, the spatial transformation matrix W r is used.
In a further embodiment of the present application, a dual permutation loss function is used to optimize the vectors in the embedded layer.
Specifically, the (h, r, t) rationality score in one embodiment described above is calculated: Then use the dual permutation loss function Optimizing embedded representations of triples in knowledge graphs
In a further embodiment of the present application, the implementation of the truncation process for the loss function is as follows:
And learning the scoring function by using the loss function, setting a truncated value of the loss function, setting the result calculated by the loss function to be zero when the loss value calculated by the loss function is larger than the truncated value and real interaction occurs between the user and the article, otherwise, keeping the result calculated by the loss function.
Specifically, the loss function can be any one of a range loss function, a square loss function and a minimum binary cross entropy loss function, the scoring function is learned through the loss function, and then the value calculated by the loss function is truncated so as to achieve the optimization processing of the scoring function, thereby filtering the interaction effect brought to implicit feedback when the loss value is larger.
Taking the minimization of the binary cross entropy loss function as an example in the present embodiment, the scoring function is learned by minimizing the binary cross entropy loss functionThe training process of (2) is as follows:
Where L CE (D) represents the loss function, Representing the actual interaction situation, the tags that have been determined,Representing the probability of interaction predicted by the scoring function.
Then, the minimized binary cross entropy loss function is truncated, so that the influence of false positive interaction with the loss value larger than the truncated value can be reduced in different stages of model training, especially in the early stage:
Where L T-CE (D) represents the truncation loss function, And when the interaction behavior between the user and the article occurs.
When the loss value is larger than the cut-off value and the real interaction occurs between the user and the object, the real interaction means that the user purchases the object or watches a certain movie, the output of the most loss function is set to zero, otherwise, the false positive interaction means that the user does not purchase the object or watches a certain movie, and the like, and the result calculated by the loss function is reserved.
Obtaining the truncated value tau: dynamic adjustment (because τ needs to be changed during different training phases). From the dynamic discard rate (dropout), i.e., ε (T), we can get:
τ (T) =ε (T) =min (αt, ε max), where ε (T) represents a time-varying cut-off value, α is a coefficient, which may be a constant, αt represents a time-varying value, and min () represents a minimum value between the two.
As shown in the above embodiments, by setting different cut-off values τ in different phases of training, in particular in the initial phase, the interaction effect with a larger loss value can be filtered out automatically to the greatest extent.
In a further embodiment of the present application, the implementation of the re-weighting of the loss function is as follows: the scoring function is learned using the loss function, and weights are added to the loss function to reduce the loss value of the loss function.
Specifically, taking minimizing the binary cross entropy loss function as an example in the present embodiment, the scoring function is learned by minimizing the binary cross entropy loss functionIs a process of (2). The training process is as follows:
The weight is added to the minimized binary cross entropy loss function, so that the influence of interaction with larger loss value is automatically reduced in the whole course of model training, and meanwhile, the influence generated by interaction with smaller loss value is reserved:
L R-CE(u,i)＝ω(u,i)LCE (u, i), where L R-CE (D) represents a truncation loss function, ω (u, i) is an added weight;
wherein, In the method, in the process of the invention,When the user and the article are shown to have interaction behavior, beta is a constant in the power function, and the value range is 0 to 1; when (when)In the time-course of which the first and second contact surfaces,Less than zero, in order to ensure that the weights ω (u, i) are greater than zero, a transform form is required.
The method is obtained by the weight increasing process, and the influence of larger interaction of the loss value can be reduced to the greatest extent by adding a new weight which is far smaller than 1 to the loss function in all the training stages, particularly in the initial training stage, for example, the model training time is 500 times, the weight can be set to be small in the initial stages from 0 to 50 times, so that the calculated loss value in the initial stage is lower, different weights are set in each stage, and the influence of the larger interaction of the loss value can be automatically reduced in the whole model training process, and meanwhile, the influence caused by the interaction with smaller loss value is reserved.
In a further embodiment of the present application, the loss function is iteratively trained, so as to minimize the loss value calculated by the loss function, and simultaneously, the loss function is noise-reduced in the iterative training process, so as to obtain an optimized scoring function.
Specifically, using the loss function to learn the scoring function, the adaptive noise reduction training process can be expressed as:
Θ＝minLCE(D*)→Θ＝minLCE(denoise(D*))
To the left of the arrow, the target Θ representing the training phase of the prior art is: the loss function is minimized by iterative training. On the right side of the scissors, the target Θ representing the training phase of the embodiment of the present application is: and carrying out noise reduction processing and minimizing a loss function in the iterative training process. In the training stage, the model is subjected to noise reduction treatment so as to achieve the purpose of filtering noise.
Embodiment two:
Based on the same inventive concept, a second embodiment of the present application provides a recommendation system based on adaptive noise reduction training, for implementing the recommendation method described in the first embodiment, where the implementation of the recommendation system may be referred to the description of a part of the method embodiments, and the repetition is not described again, including,
The generating unit 20 is configured to acquire an interaction data set of the user and the article and construct a data interaction diagram of the user and the article, add attribute data of the user and the article in the data interaction diagram, generate an attribute data relationship diagram of the user and the article, and generate a knowledge diagram including the user, the article and the attribute data according to the data interaction diagram and the attribute data relationship diagram;
The aggregation unit 30 aggregates the characteristics of the nodes adjacent to each entity node in the knowledge graph by utilizing the graph convolutional neural network, and performs splicing processing on the characteristics of the entity node and the characteristics of the nodes adjacent to the entity node through the full connection layer to obtain the characteristic vector of the user and the object;
An optimizing unit 40, configured to learn the scoring function through the loss function in the neural model training stage, and perform truncation or re-weighting on the loss function, so as to perform optimization on the scoring function;
and the recommendation unit 50 is used for performing inner product calculation on the feature vectors of the user and the article by using the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommendation result.
The recommendation system of the second embodiment performs truncation processing or re-weighting processing on the loss function in the training stage, and can automatically perform noise reduction processing in the training stage of the model by using any processing method. Compared with the prior art, the recommendation system has the advantages that: the prior recommendation model does not consider that the noise problem accompanied by implicit feedback is processed in the training stage; the application optimizes the scoring function by carrying out self-adaptive noise reduction treatment on the loss function, thereby greatly reducing the noise influence of false positive interaction on the model, reducing the workload of manually screening out the false positive interaction, and relieving the premature fitting phenomenon of the model.
In yet another embodiment of the second embodiment of the present application, the generating unit 20 of the recommendation system is further configured to construct an entity alignment matrix of the user and the item according to the data interaction diagram and the attribute data relationship diagram, and use the entity alignment matrix to compose a knowledge diagram including the user, the item and the attribute data from the data interaction diagram of the user and the item and the attribute data relationship diagram of the user and the item.
In yet another embodiment of the second embodiment of the present application, the recommendation system further includes a dimension reduction unit, where the dimension reduction unit is configured to perform dimension reduction processing on the high-dimension vector in the knowledge graph to a low-dimension vector by using an embedding layer of the graph convolutional neural network before aggregation is performed by using the graph convolutional neural network.
In yet another embodiment of the second embodiment of the present application, the dimension reduction unit is further configured to optimize the vector in the embedded layer by using a dual permutation loss function.
In yet another embodiment of the second embodiment of the present application, the optimizing unit 40 includes a first optimizing unit, where the first optimizing unit is configured to learn the scoring function by using the loss function, set a cutoff value of the loss function, set a result calculated by the loss function to zero when the loss value calculated by the loss function is greater than the cutoff value and the real interaction occurs between the user and the object, and otherwise, keep the result calculated by the loss function.
In yet another embodiment of the second embodiment of the present application, the optimizing unit 40 includes a second optimizing unit for learning the scoring function by using the loss function, and adding a weight to the loss function to reduce the loss value of the loss function.
In yet another embodiment of the second embodiment of the present application, the first optimization unit and the second optimization unit further include an objective function unit, where the objective function unit is configured to perform iterative training on the loss function, minimize a loss value calculated by the loss function, and simultaneously reduce noise of the loss function in the iterative training process, to obtain an optimized scoring function.
Embodiment III:
Based on the same inventive concept, a third embodiment of the present application provides an electronic device, including: processor 310, communication interface 320, memory 330 and communication bus 340, wherein processor 310, communication interface 320 and memory 330 communicate with each other via communication bus 340; the memory 330 is used for storing a computer program; the processor 310 is configured to execute the program stored in the memory 330, and implement the following steps: acquiring an interaction data set of a user and an article, constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article in the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram; the method comprises the steps that characteristics of nodes adjacent to each entity node in a knowledge graph are aggregated by utilizing a graph convolutional neural network, and the characteristics of the entity node and the characteristics of the nodes adjacent to the entity node are spliced through a full connection layer, so that characteristic vectors of users and articles are obtained; in the training stage of the nerve model, the scoring function is learned through the loss function, and the loss function is truncated or re-weighted, so that the scoring function is optimized; and carrying out inner product calculation on the feature vectors of the user and the article by using the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommendation result.
Embodiment four:
based on the same conception, the fourth embodiment of the present application provides a computer readable storage medium storing a computer program, where the computer program when executed by a processor implements the recommendation method based on adaptive noise reduction training described in the first embodiment.
Specifically, the instructions, when executed, perform the operations of the recommendation method included in the first embodiment. The description of the recommended method will not be described in detail here.
The foregoing description of the embodiments has been provided for the purpose of illustrating the general principles of the invention, and is not meant to limit the scope of the invention, but to limit the invention to the particular embodiments, and any modifications, equivalents, improvements, etc. that fall within the spirit and principles of the invention are intended to be included within the scope of the invention.
Claims (6)
1. The recommendation method based on the adaptive noise reduction training is characterized by comprising the following steps of:
Acquiring an interaction data set of a user and an article, constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article in the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram; generating a knowledge graph comprising user, article and attribute data according to the data interaction graph and the attribute data relation graph, wherein the knowledge graph comprises: constructing an entity alignment matrix of a user and an article according to the data interaction diagram and the attribute data relation diagram, and forming a knowledge diagram comprising the user, the article and attribute data by utilizing the entity alignment matrix;
Aggregating the characteristics of nodes adjacent to each entity node in the knowledge graph by using a graph convolution neural network, and performing splicing processing on the characteristics of the entity node and the characteristics of the nodes adjacent to the entity node through a full-connection layer to obtain the characteristic vector of the user and the object; before aggregation is carried out by using a graph convolution neural network, carrying out dimension reduction processing on high-dimensional vectors in the knowledge graph to obtain low-dimensional vectors by using an embedding layer of the graph convolution neural network, and optimizing the vectors in the embedding layer by using a dual-permutation loss function;
in the training stage of the nerve model, learning a scoring function through a loss function, and carrying out truncation or re-weighting treatment on the loss function so as to optimize the scoring function; the specific implementation of the truncation processing of the loss function is as follows: learning a scoring function by using a loss function, setting a truncated value of the loss function, setting a result calculated by the loss function to be zero when the loss value calculated by the loss function is larger than the truncated value and real interaction occurs between a user and an article, otherwise, reserving the result calculated by the loss function;
And carrying out inner product calculation on the feature vectors of the user and the article by using the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommendation result.
2. The recommendation method based on adaptive noise reduction training according to claim 1, wherein the re-weighting of the loss function is implemented as follows: the scoring function is learned using the loss function, and weights are added to the loss function to reduce the loss value of the loss function.
3. The recommendation method based on adaptive noise reduction training according to claim 1 or 2, wherein the loss function is iteratively trained to minimize a loss value calculated by the loss function, and noise reduction is performed on the loss function in the iterative training process to obtain an optimized scoring function.
4. A recommendation system based on adaptive noise reduction training is characterized by comprising,
The generating unit is used for acquiring an interaction data set of the user and the article and constructing a data interaction diagram of the user and the article, adding attribute data of the user and the article into the data interaction diagram, generating an attribute data relation diagram of the user and the article, and generating a knowledge diagram comprising the user, the article and the attribute data according to the data interaction diagram and the attribute data relation diagram; generating a knowledge graph comprising user, article and attribute data according to the data interaction graph and the attribute data relation graph, wherein the knowledge graph comprises: constructing an entity alignment matrix of a user and an article according to the data interaction diagram and the attribute data relation diagram, and forming a knowledge diagram comprising the user, the article and attribute data by utilizing the entity alignment matrix;
The aggregation unit is used for aggregating the characteristics of the nodes adjacent to each entity node in the knowledge graph by using the graph convolution neural network, and performing splicing processing on the characteristics of the entity nodes and the characteristics of the nodes adjacent to the entity nodes through the full connection layer to obtain the characteristic vectors of the user and the article; before aggregation is carried out by using a graph convolution neural network, carrying out dimension reduction processing on high-dimensional vectors in the knowledge graph to obtain low-dimensional vectors by using an embedding layer of the graph convolution neural network, and optimizing the vectors in the embedding layer by using a dual-permutation loss function;
The optimizing unit is used for learning the scoring function through the loss function in the training stage of the nerve model, and carrying out truncation or re-weighting treatment on the loss function so as to optimize the scoring function; the specific implementation of the truncation processing of the loss function is as follows: learning a scoring function by using a loss function, setting a truncated value of the loss function, setting a result calculated by the loss function to be zero when the loss value calculated by the loss function is larger than the truncated value and real interaction occurs between a user and an article, otherwise, reserving the result calculated by the loss function;
And the recommending unit is used for carrying out inner product calculation on the feature vectors of the user and the article by utilizing the scoring function after the optimization processing to obtain the matching score between the user and the article, and sequencing the matching score to generate a recommending result.
5. An electronic device, comprising: the device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are communicated with each other through the communication bus; the memory is used for storing a computer program; the processor is configured to execute a program stored in the memory, and implement the recommendation method based on adaptive noise reduction training according to any one of claims 1 to 3.
6. A computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the adaptive noise reduction training based recommendation method of any one of claims 1 to 3.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111506522.5A CN114254187B (en) | 2021-12-10 | 2021-12-10 | Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111506522.5A CN114254187B (en) | 2021-12-10 | 2021-12-10 | Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114254187A CN114254187A (en) | 2022-03-29 |
| CN114254187B true CN114254187B (en) | 2024-08-16 |
Family
ID=80794582
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111506522.5A Active CN114254187B (en) | 2021-12-10 | 2021-12-10 | Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114254187B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115525836B (en) * | 2022-08-12 | 2025-06-27 | 西安邮电大学 | A graph neural network recommendation method and system based on self-supervision |
| CN118585926B (en) * | 2024-08-07 | 2024-12-03 | 广东电网有限责任公司广州供电局 | Equipment operation efficiency evaluation method and device, storage medium and computer equipment |
Family Cites Families (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108920641B (en) * | 2018-07-02 | 2021-05-11 | 北京理工大学 | Information fusion personalized recommendation method |
| US11373233B2 (en) * | 2019-02-01 | 2022-06-28 | Target Brands, Inc. | Item recommendations using convolutions on weighted graphs |
| CN110737730B (en) * | 2019-10-21 | 2024-03-26 | 腾讯科技（深圳）有限公司 | User classification method, device, equipment and storage medium based on unsupervised learning |
| CN111222054B (en) * | 2020-01-03 | 2020-12-11 | 中国计量大学 | A Conversational Social Recommendation Method Based on Contextual Neighbor Relationship Modeling |
| CN111523051A (en) * | 2020-04-24 | 2020-08-11 | 山东师范大学 | A method and system for social interest recommendation based on graph convolution matrix decomposition |
| CN112084428B (en) * | 2020-09-17 | 2024-02-02 | 辽宁工程技术大学 | Collaborative filtering recommendation method based on coupling network embedding and knowledge graph |
| CN112417313A (en) * | 2020-11-24 | 2021-02-26 | 云南大学 | Model hybrid recommendation method based on knowledge graph convolutional network |
| CN112507224B (en) * | 2020-12-11 | 2021-09-28 | 南京大学 | Service recommendation method of man-machine object fusion system based on heterogeneous network representation learning |
| CN113434761B (en) * | 2021-06-25 | 2024-02-02 | 平安科技（深圳）有限公司 | Recommendation model training method, device, computer equipment and storage medium |
-
2021
- 2021-12-10 CN CN202111506522.5A patent/CN114254187B/en active Active
Non-Patent Citations (1)
| Title |
|---|
| 基于知识增强图卷积网络的推荐方法研究;杨敬明;万方数据;20230630;1-50 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114254187A (en) | 2022-03-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Xin et al. | CFM: Convolutional factorization machines for context-aware recommendation. | |
| CN112000819B (en) | Multimedia resource recommendation method and device, electronic equipment and storage medium | |
| CN111127142B (en) | An item recommendation method based on generalized neural attention | |
| CN110781409A (en) | Article recommendation method based on collaborative filtering | |
| CN111695965B (en) | Product screening method, system and equipment based on graphic neural network | |
| CN114254187B (en) | Recommendation method, system, electronic equipment and medium based on self-adaptive noise reduction training | |
| WO2015188349A1 (en) | Recommending of an item to a user | |
| CN107103049A (en) | A kind of recommendation method and the network equipment | |
| CN113610610B (en) | Conversational recommendation method and system based on graph neural network and review similarity | |
| CN108595533B (en) | Article recommendation method based on collaborative filtering, storage medium and server | |
| CN111967892A (en) | Information recommendation method and device | |
| CN113868537B (en) | A recommendation method based on multi-action conversation graph fusion | |
| US11651255B2 (en) | Method and apparatus for object preference prediction, and computer readable medium | |
| CN116049536A (en) | A kind of recommended method and related device | |
| CN114780753A (en) | Dialogue recommendation method, device and equipment based on knowledge graph and storage medium | |
| CN112381620A (en) | Product recommendation method and device, electronic equipment and computer-readable storage medium | |
| CN118214927A (en) | Video recommendation method and device and computer equipment | |
| Seng et al. | Item-based collaborative memory networks for recommendation | |
| CN117171449B (en) | Recommendation method based on graph neural network | |
| CN113807905B (en) | Item recommendation method and device, computer storage medium, and electronic device | |
| CN118113939A (en) | Recommendation method, system, equipment and storage medium based on causal reasoning | |
| CN112036987A (en) | Method and device for determining recommended commodities | |
| CN117493666A (en) | Article cold start recommendation method and article cold start recommendation model training method | |
| CN115730143A (en) | Recommendation system, method, terminal and medium based on task alignment meta learning and augmentation graph | |
| Gadekula et al. | Improved pearson similarity for collaborative filtering recommendation system |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |