CN115630223A - Service recommendation method and system based on multi-model fusion - Google Patents
Service recommendation method and system based on multi-model fusion Download PDFInfo
- Publication number
- CN115630223A CN115630223A CN202211189701.5A CN202211189701A CN115630223A CN 115630223 A CN115630223 A CN 115630223A CN 202211189701 A CN202211189701 A CN 202211189701A CN 115630223 A CN115630223 A CN 115630223A
- Authority
- CN
- China
- Prior art keywords
- service
- application
- text
- model
- recommendation method
- 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
- 238000000034 method Methods 0.000 title claims abstract description 39
- 230000004927 fusion Effects 0.000 title claims abstract description 25
- 230000003993 interaction Effects 0.000 claims abstract description 44
- 239000013598 vector Substances 0.000 claims description 49
- 230000006870 function Effects 0.000 claims description 11
- 238000004590 computer program Methods 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 8
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000010586 diagram Methods 0.000 claims description 5
- 238000001914 filtration Methods 0.000 claims description 5
- 230000007246 mechanism Effects 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 239000011159 matrix material Substances 0.000 claims description 3
- 238000000844 transformation Methods 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims description 2
- 238000003062 neural network model Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 239000000284 extract Substances 0.000 claims 1
- 239000010410 layer Substances 0.000 description 47
- 238000004364 calculation method Methods 0.000 description 7
- 238000012549 training Methods 0.000 description 7
- 230000000644 propagated effect Effects 0.000 description 6
- 206010054949 Metaplasia Diseases 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000015689 metaplastic ossification Effects 0.000 description 5
- 238000013459 approach Methods 0.000 description 4
- 238000011161 development Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000014509 gene expression Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 229910052731 fluorine Inorganic materials 0.000 description 1
- 125000001153 fluoro group Chemical group F* 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005295 random walk Methods 0.000 description 1
- 230000011664 signaling Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
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
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Machine Translation (AREA)
Abstract
The invention relates to a service recommendation method and a service recommendation system based on multi-model fusion, and belongs to the technical field of service recommendation. Text information, category information, text information and category information of the service, historical service call records and other contents of the application program to be created are comprehensively considered during service recommendation. The method comprises two models, wherein the preferences of the application program to the service are captured from the perspective of user requirements and historical interaction respectively, and the framework further fuses the two preferences to predict a final service recommendation list.
Description
Technical Field
The invention relates to a service recommendation method and a service recommendation system based on multi-model fusion, and belongs to the technical field of service recommendation.
Background
With the rapid development of the Internet, a service recommendation algorithm has become a mainstream paradigm for developing Web service-based applications. Developers can focus more on meeting the unique development needs of an application than writing code from scratch with the most advanced techniques, greatly freeing up the creativity of the developers. Unfortunately, the rapid growth in the number of services available in a Web application programming interface directory (e.g., the programable Web) presents a significant challenge to selecting the correct service for an application. Therefore, how to recommend appropriate services to develop new applications remains one of the key technologies.
At present, scholars at home and abroad make a lot of valuable research works in the aspect of service recommendation. One is to explore similarity matching between application descriptions and services, and generally use the functional descriptions of the services to measure the similarity, thereby completing service recommendations. For example, gu et al use a topic model to construct a semantic service package repository, and propose a service package recommendation model based on combined semantics. However, the performance of the topic model-based approach is unsatisfactory due to small data volume, noisy data, neglecting word order, and so on. And the second is the research trend in recent years of adopting deep learning algorithm to recommend. Unlike the conventional method, the recommendation method based on deep learning can automatically learn representative features from raw data. Wang et al propose an unsupervised service recommendation method based on deep random walk on a knowledge graph, which constructs the knowledge graph by using the relationship between service and mashup, realizes the implicit embedded representation of each node by using a Skip-Gram model, and calculates the correlation between the mashup node and the service node to obtain a service recommendation list.
In designing a service recommendation model, the method based on collaborative filtering and content analysis is still weak in feature fusion and utilization. The CF-based approach can share experience similar to applications, and can achieve certain effects in cases of sparse data, but does not adequately address the match between service and application requirements. In contrast, content-based approaches rely heavily on descriptive documents to estimate the correlation between demand and service. However, if sufficient information is not provided in the textual requirements, the suggested results will not be satisfactory.
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a service recommendation method and a recommendation system based on multi-model fusion, which comprise a semantic interaction model and a historical interaction model. In general, natural texts are used for describing development requirements of application programs, and in consideration of great progress of deep learning technology in natural language processing, semantic features of semantic requirements are extracted by using an NLP (non-line-of-sight) model and are expressed as vector representations. Because it is difficult to find mathematical functions that describe these interactions, graphical neural network techniques are employed in this framework to learn content interactions and history interactions between these applications and services. Finally, we integrate the features learned from content interaction and history interaction to make more accurate recommendations.
In order to achieve the above object, a first aspect of the present invention provides a service recommendation method based on multi-model fusion, including the following steps:
step (1), preprocessing a text: the description of the application to be created is denoted as a des Performing text filtering, abbreviation replacing and word metaplasia processing;
and (2) extracting text description representation: extracting text and category description by using a BERT model, converting the text description of the application program to be created, the category of the application program to be created, the text description of the service and the category of the service into characteristic vectors through the BERT model, and respectively representing the characteristic vectors as vb m ，vb mt ，vb s And vb st ；
Integrating the representation of the text features and the category features:
and (3.1) calculating the final text feature vector vs of the application program by using an attention mechanism m (referring to the final text feature vector of the application to be created) and the final text feature vector of the service vs s ；
Step (3.2), measuring the similarity between the final text feature vector of the application program and the final text feature vector of the service through element multiplication;
step (3.3), capturing text interaction between an application program to be created and a service by utilizing an MLP layer, and selecting a parameter correction linear unit (PReLU) as an activation function;
step (4), creating an application-service matrix through the application service call record, and constructing an application service network graph, wherein G = (V, E), V represents a node set, and V = A ≧ U S, and A = { a = } S 1 ，a 2 ...a p Denotes a set of applications, S = { S = } 1 ，s 2 ...s q Denotes a service set, E denotes an edge set,
if a call occurs between application a and service s, they are connected in graph G, forming E m，s Collecting edges;
bayesian Personalized Ranking (BPR) loss is used as the dominant loss;
calculating the similarity of the application program and the service;
and (5) combining the semantic interaction model and the historical interaction model, and expressing as follows:
wherein alpha is expressed as the importance of the semantic interaction model,
for the similarity between the existing application a and the service s,
is the similarity between the application and the service to be created.
The step (3.2) is specifically as follows:
given the final text feature vector vs of the application m And service final text feature vector vs s The two are connected, and the specific formula is shown as follows:
in step (3.3), the learning process is as follows:
the step (4) is specifically as follows:
step (4.1) the representation of the existing application and the service that has been called is obtained by a LightGCN that captures the non-linear relationship between the existing application and the service that has been called; carrying out information dissemination on the application program service network diagram;
step (4.2) connecting the characteristics of all K layers to combine the information received from the neighbors with different depths to obtain the final application program structural characteristics and service structural characteristics;
and (4.3) calculating by using the inner product as follows:
step (4.4). Bayesian Personalized Ranking (BPR) loss is used as the primary loss.
The step (4.1) is specifically as follows: the representation of the existing application and the invoked service is obtained through a LightGCN that captures the non-linear relationship between the existing application and the invoked service; information propagation is carried out on the application service network diagram, and the information propagation of the k layer is represented as follows:
wherein e is a (k) Denoted as k-th layer information propagation to applications a, e s (k) Denoted as k-th layer information propagation to service s, N a Denoted as neighbours of application a or neighbours of neighbours, N s Represented as a neighbor of service s or a neighbor of a neighbor, removes self-joins from the application service graph and removes non-linear transformations from the information propagation function.
The step (4.2) is specifically as follows: connecting features of all K layers to combine information received from neighbors at different depths, final application structural feature e a And service structural features e s Can be expressed as follows:
step (4.4) Bayesian Personalized Ranking (BPR) loss is used as the main loss specifically:
wherein Q = { (a, S, S ') | a ∈ A, S, S' ∈ S, x as ＝1，x as′ =0, and σ (·) is a sigmoid function.
The second aspect of the invention provides a service recommendation system based on multi-model fusion, which comprises a semantic interaction module, a history record interaction module and a multi-model fusion module;
the semantic interaction module (namely a semantic interaction model) utilizes text modeling and feature representation technology to model the description and the category of the application program to be created and the description and the category of the service, and captures the interaction between the application program to be created and the service from the content perspective;
the historical record interaction module (namely a historical interaction model) obtains the existing application program and the called service according to the historical calling record (namely the application program service calling record, also called as the historical interaction record), and models the structural information through the graph neural network model to further obtain the structural characteristics (namely the structural characteristics of the existing application program) and the structural characteristics of the service, so as to obtain the similarity between the existing application program and the service;
multi-model fusion modules are based on fusion features (i.e.
And
) And effectively fusing the semantic interaction model and the historical interaction model.
In addition, the method also comprises a text preprocessing module for realizing the step (1).
The semantic interaction module realizes the step (2) and the step (3); the history record interaction module realizes the step (4); and (5) realizing the step (5) by the multi-model fusion module.
A third aspect of the invention is to provide a processor for executing a computer program, the computer program executing the aforementioned recommendation method.
The invention also provides a terminal, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to realize the recommendation method.
It is a further object of the invention to provide a computer-readable medium, on which a computer program is stored, which computer program is executed by a processor and which can implement the aforementioned recommendation method.
Has the beneficial effects that: the method and the system provided by the invention comprehensively consider the contents of text information, category information, text information and category information of service, historical service call records and the like of the application program to be created when service recommendation is carried out. The method comprises two models, the preference of the application program to the service is captured from the perspective of user requirements and historical interaction, and the framework further fuses the two preferences to predict a final service recommendation list. The accuracy of the service recommendation method based on multi-model fusion is proved through extensive experiments under real data captured from the programable Web website.
Drawings
FIG. 1 is a block diagram of a service recommendation method based on multi-model fusion.
Detailed Description
The following describes a service recommendation method and a recommendation system based on multi-model fusion provided by the present invention. For convenience of description, the associated symbols are defined as follows:
TABLE 1 symbol definitions
When creating an application program, a developer inputs application program requirements or description (application program requirements are application program description) according to the general technical requirements of creating the application program, the description adopts natural language and is marked as a des ，a des The system is composed of a group of phrases, sentences and even paragraphs, and is used as the development requirement of developers on similar application programs. The application requirements or descriptions constitute text, and the descriptions typically contain noise and therefore cannot be used directly for model learning. This application that has been required or described is referred to as the application to be created. For example, an application requires or describes: i want to develop a travel system that can book airline tickets and hotels.
Preprocessing a text, wherein in order to keep important information in the text of an application program to be created, preprocessing processes such as text filtering, abbreviation replacement and word metaplasia need to be carried out to obtain the preprocessed text;
text filtering: invalid words such as labels, punctuation, non-characters, etc. are usually meaningless and are filtered out by regular expressions. In addition, the stop word list is used to delete stop words in the text.
Abbreviation substitution: replacing the english abbreviation with the complete spelling; english abbreviations, such as don't, i'm.
And (3) word metaplasia: lexical metaplasia is the process of transforming a word into its basic form. The method of word metaplasia is based on a built-in deformation function of WorldNet. For example, the word "used" becomes used.
And (3) realizing the step (1) by a text preprocessing module.
And (2) extracting text description expression: extracting texts and category descriptions from the preprocessed texts by using a natural language technology BERT model, and converting text requirements (text descriptions), categories of applications to be created, text descriptions of services and categories of services of the applications to be created into text feature vectors respectively represented as vb through the BERT model m (application text description feature vector), vb mt (application class feature vector), vb s (service text description feature vector) and vb st (service class feature vector);
respectively extracting text description and category description from the preprocessed application program requirement or description through a BERT model, and converting the extracted text requirement (namely text description) of the application program into an application program text description feature vector vb through the BERT model m (ii) a Converting the extracted application category description into an application category feature vector vb through a BERT model mt 。
In some embodiments, through the BERT model, the category description of the application to be created can be extracted by combining the text description of the application to be created and the existing category descriptions of many applications.
For example, the service in step (2) may be any one of all services in a progrmmablable web site. Programma pairs using BERT modelExtracting text and category description from a certain service in the bleWeb website, more specifically, extracting the text description and the category description of the service from the webpage by using a BERT model, wherein each service is in the webpage; converting the extracted text description of a certain service into a service text description feature vector vb through a BERT model s (ii) a Converting the extracted class description of a certain service into a service class feature vector vb through a BERT model st . The service is crawled from the existing web pages, for example, a Baidu map is a service and can be reused in each web page. There is a certain probability that any service matches the application to be created.
Step (3) consider that the text requirements and class characteristics of the application have different effects on the overall text feature generation, which means that we should assign different weights to the representation of the text requirements and class characteristics in question. An attention-based approach is devised to integrate the representation of textual requirements and category characteristics; the weights are assigned according to an attention algorithm, and the attention algorithm can be used for determining the weight assigned to the text requirement and the weight assigned to the category.
Step (3.1) of inputting text characteristics and category characteristics of application programs, wherein M = [ vb ] m ，vb mt ]Calculating the final text feature vector vs of the application program by using an attention mechanism m . The specific process is as follows:
calculating the attention distribution of the text feature and the category feature of the application program, and enabling Key = Value = M = q to give the attention distribution: a is a i ＝softmax(s(key i ，q))＝softmax(s(M i Q)), where s (M) i Q) is M i The point-of-attention scoring mechanism of (1), here using a scaled dot product model, the formula is as follows:
key, value, q are three designated inputs to the attention algorithm, where M i e.M, i =0 or 1,M i For text or category features of the application to be created, i.e. M 0 For the text features (i.e. vb) of the application to be created m )，M 1 For class characteristics (i.e. vb) of the application to be created mt )，a i Attention distribution for application text features or category features of an application to be created, i.e. a 0 Attention distribution for text features of an application to be created, a 1 Attention distribution for class features of the application to be created; m i T Is M i The transposing of (1).
Wherein,
represents a scaling factor, here taken as the length of a text feature (i.e., the application text description feature vector vb) m Length of the application), and finally calculating the final text feature vector vs of the application m The formula is as follows:
input service text feature and category feature S = [ vb ] s ，vb st ]Computing the final text feature vector vs for the service using the attention mechanism s . The calculation mode and the final text feature vector vs of the application program m The same way of calculation.
Step (3.2) in general, the similarity between two vectors can be measured by elemental multiplication. Given the final text feature vector vs of the application m And service final text feature vector vs s We link the two together, as shown in the following formula:
vs ms and the final text feature vector of the application program and the final text feature vector of the service are connected.
Step (3.3) this document utilizes the MLP (Multi-layer perceptron neural network) layer to capture the application to be createdAnd a certain service, furthermore we choose a parameter correcting linear unit (PReLU) as the activation function, since it can improve the model fitting with almost zero additional computational cost and little risk of overfitting. Input vs ms And initialization bias b = { b = { b } 1 ，b 2 ，...b z The learning process is as follows:
wherein W z And b z Denotes the connection weights and offset vectors between z-layers and (z-1) layers. The advantage of MLP is that it can learn the interaction characteristics at different abstraction levels. As the number of layers increases, the receptive field of each neuron becomes larger relative to the previous layer, so it can provide global semantics (global interaction) and abstract details, which are difficult to do in shallow and linear operations;
where z is obtained by experiment, where z is 3, layer 1 is the input layer, and the input is vs ms The 2 nd layer is a hidden layer, and the 3 rd layer is an output layer. W is a group of 1 Is vs ms B1 is vs ms The offset vector of (a); w 2 Is the connection weight between layer 2 and layer 1; b2 is the offset vector between layer 2 and layer 1. W 3 The connection weight between the 3 rd layer and the 2 nd layer is given; b3 is the offset vector between layer 3 and layer 2.
W (connection weight) maps the input variables into a new dimensional space. The presence of b provides the mapped data with translation capability, referred to as offset.
Input is vs ms The single-layer learning process is (W) 1 (vs ms ))+b 1 ). The weight and offset refer to the weight and offset between layers.
Step (4), creating an application-service matrix through the application service call record, and constructing an application service network graph, wherein G = (V, E), V represents a node set, and V = A ≧ U S, and A = { a = } S 1 ，a 2 ...a p Denotes a set of applications, S = { S = } 1 ，s 2 ...s q Denotes a service set, E denotes an edge set. a is 1 For the 1 st existing application, a 2 For the 2 nd existing application, a p For the pth existing application, there are a total of p existing applications. s 1 For the 1 st service that has been called, s 2 For the 2 nd called service, s q For the q-th called service, the number of called services is q in total.
If a call occurs between application a and service s, they are connected in graph G, forming E m，s And (5) edge collection. The service s is the same service as the service in step (2).
e h，j Denotes a h And s j Edge in between, if present, e h，j =1, otherwise 0.a is h Represents the h-th existing application, s j Indicating the jth service that has been called. h =1, 2.. P; j =1, 2.. Q. When the h existing application program calls the jth called service, the h existing application program is connected with the jth called service, and the connection line is a h And s j Edge in between, at this time e h，j =1; when the h existing application program does not call the j called service, the h existing application program is not connected with the j called service and does not form an edge, and at the moment e h，j ＝0。
The pairwise combinations (calls occurring) between service and application are l, so h, j =1, 2.. L, i.e. there are l connecting lines, i.e. l edges, in graph G.
The service in this step is the service that has been called in the application service call record. The application in this step is an existing application, and may be all existing applications in the programammable web site.
Step (4.1) a representation of existing applications and services that have been invoked may be obtained by the LightGCN, which may capture the non-linear relationship between the applications and the services. Specifically, we perform information propagation on the application service network graph, and the information propagation of the k-th layer can be expressed as:
wherein e is a (k) Denoted as k-th layer information propagation to applications a, e s (k) Denoted as layer k information propagation to service s, application a is a 1 ，a 2 ...a p Of the application, service s being s 1 ，s 2 ...s q A service in (2) that has been called by application a, N a Denoted as neighbours of application a or neighbours of neighbours, N s A neighbor denoted as a neighbor of service s or a neighbor of a neighbor; we remove self-joins from the application service network graph and remove non-linear transformations from the information propagation function; the starting point and the ending point of the self-connecting finger point are the same. The nonlinear transformation refers to a nonlinear activation function, and the characteristic transformation of the application program characteristic information and the service characteristic information is cancelled.
According to the experimental results, k =0, 1, 2 here. Where k =1, 2, e can be used s (k) 、e a (k) And calculating by using a formula. When k =0, e s (0) The structural characteristics of the service s itself; e.g. of the type a (0) Is a structural feature of the application a itself. e.g. of a cylinder a (0) 、e s (0) Feature vectors (defined here as structural feature vectors for the purpose of distinguishing from feature vectors in the semantic interaction model) automatically obtained after LightGCN is input for the application a and the service s, respectively. In each layer, each node combines its neighborsTo obtain a new embedding. LightGCN uses 2 layers to accomplish the embedded training, stacking more layers means that information from a given node can get information from nodes further away from the node, so that higher order graph structures can be captured as needed. If the node represents the 0 th layer, the 1 st layer is the neighbor node of the node, and the 2 nd layer is the neighbor node of the neighbor, the nodes are stacked by using the formula, and the node vector of the node is updated.
If a service s is called by an existing application a, the application a is connected to the service s, the neighbor of the application a is the service s, if the service s is called by another existing application a n When called, the neighbors of the application a's neighbors are application a n ，a n E.g. A. Due to e a (k) Indicating that the k-th layer information is propagated to application a, assuming k is 1,e a (k) Indicating that layer 1 information (service s) is propagated to application a; let k be 2, which means layer 2 information (application a) n ) To application a.
For the same reason, because e s (k) Indicating that the k-th layer information is propagated to the service s, and assuming that k is 1, indicating that the 1-th layer information (application program a) is propagated to the service s; let k be 2, meaning layer 2 information (service s) t ) To the service s. s t I.e. another service called by application a, s t ∈S；
When k =1, at e a (k) And e s (k) In the calculation formula, N s As application programs a, N a Is service s, i.e. application a and service s are neighbors to each other in layer 1.
When k =2, at e a (k) And e s (k) In the calculation formula, N s As a service s t ，N a For application a n I.e. the neighbours of application a in layer 2 are another application and the neighbours of service s are another service.
in layer 2, e s (2) Being neighbors of a neighbour s t A structural feature vector propagated to service s information; e.g. of a cylinder a (2) Is a neighbor of a neighbor (a) n ) Structural feature vectors that are propagated to application a information.
If there is no neighbor to a service or an application, the corresponding e s (2) Or e a (2) Is 0.
Step (4.2) connect all K-layer features to combine information received from neighbors of different depths, the final application structural feature and service structural feature can be expressed as follows:
according to the experimental results, K is 2.
Step (4.3) to obtain the recommended final prediction, we first perform a calculation using the inner product, as follows:
Step (4.4) the conventional Bayesian Personalized Ranking (BPR) penalty is used as the dominant penalty, which can be expressed as follows:
wherein Q = { (a, S, S ') | a ∈ A, S, S' ∈ S, x as ＝1，x as′ =0, σ (·) is a sigmoid function.
Representing the similarity between application a and service s,
represented is the similarity between application a and service s'. x is a radical of a fluorine atom as Representing application a calling a service s, x as′ Representing that application a has not interacted with service s' in the history. Any service in the service set S that has not been called by the application a can be regarded as S'.
Calculated by substituting s' into steps 4.1-4.3.
Step (4.4) is used during training of the feature vectors (i.e. the training process of steps 4.1-4.3), the difference between positive and negative samples is made as large as possible, and training can be terminated when training is performed for a certain number of times (e.g. 1000 times) or when L is reached BPR If the value is less than a fixed value (for example, 0.000001), the training is finished. To end training e a 、e s Is calculated to obtain
Step (5) in order to improve the accuracy of recommendation, we use a linear framework to combine the semantic interaction model and the historical interaction model, which can be expressed as follows:
wherein, α represents the importance of the semantic interaction model, and is obtained by experiments, and is 0.9.y is as Indicates the general purposeUsing program and service similarities.
At the time of calculating final y as When the utility model is used, the water is discharged,
services corresponding in computing and
the corresponding services are the same. Aiming at each service, combining historical interaction records with the application program to be created to obtain corresponding y as The value is obtained. I.e. each service, has a corresponding y as Value, finally get y of several services as A list of values. If a certain service corresponds to y as Larger values are more recommended. Final y according to different services as The size of the value lists the recommendation list from large to small.
If a service has not been called by an existing application, then y as At the time of calculation, only calculation
Not calculating
The foregoing illustrates and describes the principles, general features, and advantages of the present invention. It should be understood by those skilled in the art that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by using equivalent alternatives or equivalent variations fall within the scope of the present invention.
Claims (10)
1. The service recommendation method based on multi-model fusion is characterized by comprising the following steps:
step (1), preprocessing a text: the description of the application to be created is denoted as a des Performing text filtering, abbreviation replacement and lemmatization processing;
and (2) extracting text description representation: extracting text and category description by using a BERT model, and creating the application program to be created by the BERT modelThe text description, the category of the application to be created, the text description of the service and the category of the service are converted into feature vectors, which are respectively represented as vb m ，vb mt ，vb s And vb st ；
Integrating the representation of the text features and the category features:
and (3.1) calculating the final text feature vector vs of the application program by using an attention mechanism m And final text feature vector vs of service s ；
Step (3.2), measuring the similarity between the final text feature vector of the application program and the final text feature vector of the service through element multiplication;
step (3.3), capturing text interaction between an application program to be created and a service by utilizing an MLP layer, and selecting a parameter correction linear unit (PReLU) as an activation function;
step (4), creating an application-service matrix through the application service call record, and constructing an application service network graph, wherein G = (V, E), V represents a node set, and V = A ≧ U S, and A = { a = } S 1 ，a 2 ...a p Denotes a set of applications, S = { S = } 1 ，s 2 ...s q Denotes a service set, E denotes an edge set,
if a call occurs between application a and service s, they are connected in graph G, forming E m，s Collecting edges;
bayesian Personalized Ranking (BPR) loss is used as the dominant loss;
calculating the similarity of the application program and the service;
and (5) combining the semantic interaction model and the historical interaction model, and expressing as follows:
wherein alpha is expressed as the importance of the semantic interaction model,
for the similarity between the existing application a and the service s,
is the similarity between the application and the service to be created.
2. The multi-model fusion-based service recommendation method according to claim 1,
the step (3.2) is specifically as follows:
given the final text feature vector vs of the application m And service final text feature vector vs s And connecting the two, specifically as shown in the following formula:
in step (3.3), the learning process is as follows:
3. the service recommendation method based on multi-model fusion according to claim 1, wherein the step (4) is specifically as follows:
step (4.1) the representation of the existing application and the service that has been called is obtained by a LightGCN that captures the non-linear relationship between the existing application and the service that has been called; carrying out information propagation on the application program service network diagram;
step (4.2) connecting the characteristics of all K layers to combine the information received from the neighbors of different depths to obtain the final application program structural characteristics and service structural characteristics;
and (4.3) calculating by using the inner product as follows:
step (4.4). Bayesian Personalized Ranking (BPR) loss is used as the primary loss.
4. The multi-model fusion-based service recommendation method according to claim 1,
the step (4.1) is specifically as follows: the representation of the existing application and the invoked service is obtained through a LightGCN that captures the non-linear relationship between the existing application and the invoked service; information dissemination is performed on the application service network diagram, and the information dissemination of the k-th layer is expressed as follows:
wherein e is a (k) Denoted as k-th layer information propagation to applications a, e s (k) Denoted as k-th layer information propagation to service s, N a Denoted as neighbours of application a or neighbours of neighbours, N s Represented as a neighbor of service s or a neighbor of a neighbor, removes self-joins from the application service graph and removes non-linear transformations from the information propagation function.
5. The multi-model fusion-based service recommendation method according to claim 1,
the step (4.2) is specifically as follows: connecting the features of all K layers to combine the information received from neighbors of different depths, the final application structural feature and service structural feature can be expressed as follows:
6. the multi-model fusion-based service recommendation method according to claim 1,
bayesian Personalized Ranking (BPR) loss was used as the main loss specifically:
wherein Q = { (a, S, S ') | a ∈ A, S, S' ∈ S, x as ＝1，x as′ =0, σ (·) is a sigmoid function.
7. The service recommendation system based on multi-model fusion is characterized by comprising a semantic interaction module, a history record interaction module and a multi-model fusion module;
the semantic interaction module utilizes text modeling and feature representation to model text description and category of the application program to be created and text description and category of the service, and captures interaction between the application program to be created and the service from the perspective of text;
the historical interaction recording module extracts the call records of the existing application programs and services and models the structural information through a graph neural network model;
the multi-model fusion module effectively fuses the semantic interaction module and the historical interaction recording module based on fusion characteristics.
8. A processor for executing a computer program, characterized in that the computer program executes to perform the recommendation method according to any one of claims 1-6.
9. A terminal comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the recommendation method according to any one of claims 1-6 when executing the program.
10. A computer-readable medium, on which a computer program is stored, characterized in that the computer program is executed by a processor, wherein the recommendation method according to any one of claims 1-6 is implemented.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211189701.5A CN115630223A (en) | 2022-09-28 | 2022-09-28 | Service recommendation method and system based on multi-model fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211189701.5A CN115630223A (en) | 2022-09-28 | 2022-09-28 | Service recommendation method and system based on multi-model fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115630223A true CN115630223A (en) | 2023-01-20 |
Family
ID=84903742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211189701.5A Pending CN115630223A (en) | 2022-09-28 | 2022-09-28 | Service recommendation method and system based on multi-model fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115630223A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116739787A (en) * | 2023-08-11 | 2023-09-12 | 深圳市艾德网络科技发展有限公司 | Transaction recommendation method and system based on artificial intelligence |
-
2022
- 2022-09-28 CN CN202211189701.5A patent/CN115630223A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116739787A (en) * | 2023-08-11 | 2023-09-12 | 深圳市艾德网络科技发展有限公司 | Transaction recommendation method and system based on artificial intelligence |
| CN116739787B (en) * | 2023-08-11 | 2023-12-26 | 深圳市艾德网络科技发展有限公司 | Transaction recommendation method and system based on artificial intelligence |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111061856B (en) | Knowledge perception-based news recommendation method | |
| Shi et al. | Functional and contextual attention-based LSTM for service recommendation in mashup creation | |
| Liu et al. | Cross-domain sentiment aware word embeddings for review sentiment analysis | |
| CN117453921B (en) | Data information label processing method of large language model | |
| CN112347268A (en) | Text-enhanced knowledge graph joint representation learning method and device | |
| CN112417097B (en) | Multi-modal data feature extraction and association method for public opinion analysis | |
| CN112256866B (en) | Text fine-grained emotion analysis algorithm based on deep learning | |
| Zhao et al. | The study on the text classification for financial news based on partial information | |
| CN112232053B (en) | Text similarity computing system, method and storage medium based on multi-keyword pair matching | |
| CN113392209B (en) | Text clustering method based on artificial intelligence, related equipment and storage medium | |
| Yu et al. | A novel weight-oriented graph convolutional network for aspect-based sentiment analysis | |
| CN112015868A (en) | Question-answering method based on knowledge graph completion | |
| CN112115253B (en) | Depth text ordering method based on multi-view attention mechanism | |
| Chen et al. | A few-shot transfer learning approach using text-label embedding with legal attributes for law article prediction | |
| CN112800774A (en) | Entity relationship extraction method, device, medium and equipment based on attention mechanism | |
| CN115062727B (en) | Graph node classification method and system based on multi-order hypergraph convolutional network | |
| CN114692605A (en) | Keyword generation method and device fusing syntactic structure information | |
| CN110569355A (en) | Viewpoint target extraction and target emotion classification combined method and system based on word blocks | |
| CN114428850A (en) | Text retrieval matching method and system | |
| CN114579741B (en) | GCN-RN aspect emotion analysis method and system for fusing syntax information | |
| CN116737979A (en) | Context-guided multi-modal-associated image text retrieval method and system | |
| CN115630223A (en) | Service recommendation method and system based on multi-model fusion | |
| Sun et al. | Rumour detection technology based on the BiGRU_capsule network | |
| CN114881038B (en) | Chinese entity and relation extraction method and device based on span and attention mechanism | |
| CN116414988A (en) | Graph convolution aspect emotion classification method and system based on dependency relation enhancement |
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 |