CN112529637A - Service demand dynamic prediction method and system based on context awareness - Google Patents
Service demand dynamic prediction method and system based on context awareness Download PDFInfo
- Publication number
- CN112529637A CN112529637A CN202011526415.4A CN202011526415A CN112529637A CN 112529637 A CN112529637 A CN 112529637A CN 202011526415 A CN202011526415 A CN 202011526415A CN 112529637 A CN112529637 A CN 112529637A
- Authority
- CN
- China
- Prior art keywords
- service
- service demand
- user
- interaction
- scene
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 34
- 230000003993 interaction Effects 0.000 claims abstract description 88
- 230000007246 mechanism Effects 0.000 claims abstract description 45
- 239000013598 vector Substances 0.000 claims description 22
- 230000002452 interceptive effect Effects 0.000 claims description 20
- 238000003062 neural network model Methods 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 13
- 238000011176 pooling Methods 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 6
- 238000003860 storage Methods 0.000 claims description 4
- 230000000977 initiatory effect Effects 0.000 claims description 3
- BGUILKOLBPOILU-UHFFFAOYSA-N 2-(dimethylamino)ethyl 2-cyclohexyl-2-hydroxy-2-phenylacetate Chemical compound C=1C=CC=CC=1C(O)(C(=O)OCCN(C)C)C1CCCCC1 BGUILKOLBPOILU-UHFFFAOYSA-N 0.000 abstract description 31
- 238000012549 training Methods 0.000 abstract description 9
- 238000002474 experimental method Methods 0.000 description 18
- 238000004422 calculation algorithm Methods 0.000 description 15
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 12
- 238000005457 optimization Methods 0.000 description 10
- 238000011160 research Methods 0.000 description 9
- 238000011156 evaluation Methods 0.000 description 8
- 238000013528 artificial neural network Methods 0.000 description 5
- 210000002569 neuron Anatomy 0.000 description 5
- 230000004913 activation Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 238000013507 mapping Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004883 computer application Methods 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0201—Market modelling; Market analysis; Collecting market data
- G06Q30/0202—Market predictions or forecasting for commercial activities
Landscapes
- Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Engineering & Computer Science (AREA)
- Accounting & Taxation (AREA)
- Development Economics (AREA)
- Finance (AREA)
- Entrepreneurship & Innovation (AREA)
- Game Theory and Decision Science (AREA)
- Data Mining & Analysis (AREA)
- Economics (AREA)
- Marketing (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The method comprises the steps of firstly, capturing interaction relations between different scenes and service requirements in a self-adaptive manner through an interaction unit of an AMEDIN model, thereby explicitly modeling the influence of the different scenes on the service requirements; then, combining scene characteristics, interaction relations and service requirement characteristics, and acquiring influence weights of different scenes on service requirements based on an attention mechanism; and finally, training an AMEDIN model based on the user characteristics, the weighted scene characteristics and the service demand characteristics, and realizing the dynamic prediction of the service demand of context awareness based on the trained AMEDIN model.
Description
Technical Field
The disclosure relates to the technical field of computer application, in particular to a service demand dynamic prediction method and system based on context awareness.
Background
In recent years, with the rapid development and popularization of service computing, the internet of things, intelligent terminals and 5G networks, more and more users can access services with rich functions from different fields at any time and any place to complete work and daily life affairs. With the rapid increase of the number of available services on the network, it is difficult for users to quickly and timely find the services meeting their needs, which seriously affects the satisfaction of users and reduces the utilization rate of service resources. Active service recommendation gradually becomes a key technology for realizing intelligent service, and dynamic prediction of service requirements is a basis for realizing active service recommendation. How to realize dynamic prediction of service requirements has become one of the key problems to be solved urgently in the field of intelligent services.
In recent years, researchers at home and abroad have made researches for the problem and have obtained certain research results. The inventor finds that the existing research work is mostly based on collaborative filtering, support vector machine, matrix decomposition and machine learning method to realize the prediction of user service requirement; although the research work achieves better results, the existing research work generally regards the influence of different scenes on the service requirements of users as the same important, so that the model cannot fully learn the influence of different scenes on the service requirements, and the accuracy of service requirement prediction is reduced; meanwhile, the influence of the scene where the user is located on the service requirement of the user is not fully considered in the existing research work, so that the service requirement prediction precision is not high. In fact, the service requirement of the user has a close relationship with the scene where the user is located, and the scene where the user is located is an important factor for triggering the user to put forward the service requirement, so that the scene where the user is located needs to be fully considered when the user service requirement is predicted; in practical application, a user often puts forward the same service demand under different scenes, and the influence of the different scenes on the service demand is different, so that the influence weight of a plurality of different scenes on the service demand needs to be considered when predicting the service demand, and the accuracy of service demand prediction is improved.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention provides a service demand dynamic prediction method and system based on context awareness, the scheme constructs a depth interaction neural network model with enhanced attention mechanism, and the interaction relationship between different scenes and service demands is captured in a self-adaptive manner through interaction units of the network model, so that the influence of the different scenes on the service demands is modeled explicitly, and the interpretability and the accuracy of service demand prediction are improved.
According to a first aspect of the embodiments of the present disclosure, a method for dynamically predicting service demand based on context awareness is provided, including:
acquiring relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information; dynamically predicting the service requirement by utilizing a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
Furthermore, in order to obtain the interaction relationship between different scenes and service requirements, data needs to be preprocessed before being input into the interaction unit, and firstly, data with the same service requirements are divided into a group for each user; and then, extracting different scene characteristics in the same group of data, and combining the scene characteristics with the user characteristics and the service requirement characteristics to form a piece of sample data.
Further, after the interactive relation between different scene features and service requirement features is obtained, pooling is carried out on the scene features in an average pooling mode to obtain pooled scene features of a plurality of scene features, and the pooled scene features represent main scene features of a plurality of scenes initiating service requirements; and acquiring an interactive relation between the pooling scene characteristics and the service requirement characteristics through the interactive unit.
Further, when calculating the influence weight of different scenes on the service demand, the scene features and the interaction relation corresponding to the pooling features are combined into a new splicing vector, the vector output through the attention mechanism is the weighted sum of the splicing vector and the influence weight, and the vector represents the scene features with obvious influence degree on the service demand initiated by the user.
Further, after scene features having a large influence on user initiated service demands are obtained, the prediction module based on the attention mechanism enhanced deep interaction neural network model realizes prediction of the service demands.
Further, the attention mechanism enhanced deep interaction neural network model selects a cross entropy loss function to optimize the constructed model.
Furthermore, the interaction unit and the service demand prediction module are formed by a fully connected network.
According to a second aspect of the embodiments of the present disclosure, there is provided a dynamic service demand prediction system based on context awareness, including:
the data acquisition unit is configured to acquire relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information;
a service demand prediction unit configured to perform dynamic prediction of a service demand using a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
According to a third aspect of the embodiments of the present disclosure, there is provided an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the memory, where the processor implements the method for dynamically predicting service demand based on context awareness when executing the program.
According to a fourth aspect of the embodiments of the present disclosure, there is provided a non-transitory computer-readable storage medium, on which a computer program is stored, which when executed by a processor implements the method for dynamic prediction of service demand based on context awareness.
Compared with the prior art, the beneficial effect of this disclosure is:
(1) according to the scheme disclosed by the disclosure, an attention mechanism enhanced deep interaction neural network model AMEDIN is constructed, and a context-aware service demand dynamic prediction method is provided based on the AMEDIN. The method captures the interaction relation between different scenes and the service requirements, and obtains the influence weight of the different scenes on the service requirements, so that the scenes with strong relevance with the service requirements obtain higher influence weight, and the method plays a leading role in predicting the context-aware service requirements.
(2) The scheme of the present disclosure introduces an interaction unit. The interaction relation between a plurality of scenes and the service requirements can be explicitly modeled through the interaction unit, the nonlinear relation between different scenes and the service requirements can be fully captured, and the accuracy of service requirement prediction can be improved.
(3) According to the scheme, through the combination of the interaction unit and the attention mechanism, the influence weight of different scenes on the service demand is dynamically acquired, and the scene characteristics with large influence on the user service demand are mined, so that the interpretability and the precision of service demand prediction are improved.
Drawings
The accompanying drawings, which are included to provide a further understanding of the disclosure, illustrate embodiments of the disclosure and together with the description serve to explain the disclosure and are not to limit the disclosure.
Fig. 1 is a diagram of an amedn model structure according to a first embodiment of the present disclosure;
fig. 2(a) -2 (d) are graphs illustrating loss values of the amedn model on Movielens data sets according to the first embodiment of the disclosure;
fig. 3(a) -3 (d) are graphs illustrating loss values of the amedn model on the Alibaba data set according to the first embodiment of the disclosure;
fig. 4 is a graph illustrating the accuracy of different models in the Movielens data set according to the first embodiment of the disclosure;
FIG. 5 is a graph illustrating the accuracy of different models in the Alibaba data set according to the first embodiment of the disclosure;
FIG. 6 is a schematic diagram of RMSE on a Movielens dataset for different models described in the first embodiment of the present disclosure;
fig. 7 is a schematic MAE of different models on Movielens data set according to the first embodiment of the disclosure;
FIG. 8 is a schematic diagram of RMSE on an Alibaba data set for different models according to one embodiment of the present disclosure;
fig. 9 is a schematic diagram of MAE of different models on an Alibaba data set according to the first embodiment of the present disclosure.
Detailed Description
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless otherwise defined, all technical and scientific terms used in the present examples have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present disclosure. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
The first embodiment is as follows:
the embodiment aims to provide a service demand dynamic prediction method based on context awareness.
A dynamic service demand prediction method based on context awareness comprises the following steps:
acquiring relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information; dynamically predicting the service requirement by utilizing a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
In service use, the characteristic information, the scene information and the proposed service requirement information of a user can be obtained through the intelligent terminal, the Internet of things and the intelligent wearable device, so that relevant data when the user proposes a service requirement can be formed, and data support is provided for realizing context-aware service requirement prediction. For the problem of dynamic prediction of service demand of context awareness, a service usage data model is defined as shown in formula (1):
SAR＝<UF，CF，SRF> #(1)
wherein UF ═ is<uf1，...，ufk>The user characteristics are represented, and mainly comprise the gender, age, occupation, income and the like of the user. CF ═<cf1，...cfh>And representing scene characteristics, mainly comprising time, position, weather, event, companions and the like set forth by service demands. SRF ═<srf1，...，srf1>The service requirement characteristics are represented and mainly comprise the domain to which the service belongs, the service name, the service function, the service level and the like. The user characteristics, the scene characteristics and the service requirement characteristic vector have good expansibility, and related characteristics can be added or subtracted according to needs.
In order to improve the accuracy of dynamic prediction of the context-aware service demand, in this embodiment, an attention mechanism-enhanced deep interaction neural network model amedn is designed to obtain an interaction relationship between different scenes and the service demand, and further obtain the influence weight of the different scenes on the service demand. The amedn model and its main operation mechanism are introduced below.
Specifically, the specific structure of the attention mechanism enhanced deep interaction bible network model in this embodiment includes:
(1) interactive relationship acquisition based on interactive units
In order to obtain the interaction relationship between different scenes and service requirements, the data needs to be preprocessed before being input into the interaction unit. Firstly, aiming at the ith user, dividing data with the same service requirement into a group; then, different scene features (noted as different scene features in the same group of data) are extracted
) From CFiCombining with user characteristics and service requirement characteristics to form a sample data Xi，
Wherein, UFiA feature vector representing an ith user; SRFiIndicating that a user is present in a scene
Feature vectors of the proposed service requirements are set down; for the service usage history dataset, it is preprocessed as described above, resulting in an input dataset D ═ X for the model1,…,Xi,…,Xn}。
When inputting data to AMEDIN model, data X is inputiMultiple scene features in
And service requirement feature SRFiInputting the scene characteristics to an interactive unit, and acquiring each scene characteristic through the interactive unit
And service requirement feature SRFiThe interaction relationship between them. The obtaining formula of the interaction relationship is shown as formula (2):
wherein,
representing scene features for the output of interactive elements
And service requirement feature SRFiSigma denotes the ReLU (rectifier activation function) activation function, W1Representing a weight matrix, b1A vector of the offset is represented, and,
representing the vector splicing operator.
After the interactive relation between different scene characteristics and service requirement characteristics is obtained, the scene characteristics are subjected to average pooling
Pooling to obtain pooled scene features of multiple scene features
The pooled scene features represent main scene features of a plurality of scenes initiating service demands, and the interaction relation between the pooled scene features and the service demand features is obtained through the interaction unit
The measurement formula is shown in formula (3):
wherein,
representing average pooled scene features
And service requirement feature SRFiThe mutual relationship between the two, sigma denotes the ReLU activation function, W2Representing a weight matrix, b2A vector of the offset is represented, and,
representing the vector splicing operator. Based on the method, the interactive relation between the scene characteristics with the incidence relation and the service requirement characteristics and the interactive relation between the pooling scene characteristics and the service requirement characteristics can be obtained.
(2) Influence weight learning based on attention mechanism
Inspired by the human visual attention running mechanism, some scholars propose an attention mechanism. In recent years, attention mechanisms have been successfully applied in the fields of computer vision, natural language processing, and service recommendation. In the embodiment, the influence weight of different scenes on the service demand is acquired by introducing an attention mechanism, so that the model can effectively capture scene characteristics having important influence on the service demand, and the prediction capability and the interpretability of the model are improved. When calculating the influence weight of different scenes on service requirements, the scene characteristics are calculated
And the corresponding interaction relation
Interaction relation corresponding to pooled features
Splicing into a new vector, noted
As shown in equation (4):
interaction ei polHigh-order characteristics between the expressed scene and the service requirement are reflected by high-order interaction relation between the scene and the service, and the prediction precision of the model is improved in addition
Wherein,
representing the concatenation of the vectors. Is provided with
Is composed of
Feature on service demand SRFiThe impact weight of (a) is, according to the attention mechanism,
the calculation formula (2) is shown in formula (5):
where n represents the number of service demands historically placed by the user. The formula (5) includes two operations, one is composed of
And the inner product operation between vectors is represented, and the second is softmax function operation. The relation between the scene characteristics and the service requirement characteristics can be extracted through inner product operation; generating weight through the result of the standard inner product operation of the softmax function
Similarly, according to the formula (5), the scene characteristics can be obtained
Feature on service demand SRFiInfluence weight of
In acquiring each scene feature
Feature on service demand SRFiAfter the influence weight of (c), the output of the attention mechanism is shown in equation (6):
output of attention mechanism
As a spliced vector
And the weighted sum of the influence weights, the vector represents the scene characteristics with large influence on the user initiated service demand.
(3) Dynamic prediction of service demand
After scene characteristics which have large influence on user initiated service requirements are obtained, the prediction module based on the AMEDIN model realizes prediction of the service requirements. When the service demand forecasting module is trained, the input data of the model is defined as
Wherein, UFiA feature vector representing the user is generated,
representing weighted scene features derived from attention mechanisms, SRFiRepresenting a service demand characteristic; y isiLabels representing input data, yiE {0,1}, when yiWhen equal to 0, it means that the user does not have the current SRFiA corresponding service requirement; when y isiWhen 1, it indicates that the user has the current SRFiA corresponding service requirement; the learning function of the service demand prediction module is shown in equation (7):
where σ denotes the ReLU activation function, W denotes the weight matrix, IiRepresenting the input data and b representing the offset vector.
In the amedn model, a cross entropy loss function (CrossEntropyLoss) is selected to optimize the constructed model, and the cross entropy loss function is shown as formula (8):
where M represents the data set, N is the number of data samples, I represents the input to the model, y represents the actual label of the input data,
representing a predicted label. In view of the fact that the Adam optimization algorithm has a good optimization effect, in the service demand prediction model, Adam is selected as the optimization algorithm of the amedn model in the embodiment. In each update of the model, the Adam algorithm calculates and corrects the first moment deviation and the second moment deviation of the model parameters, and then moves along the negative direction of the parameter gradient until the model is updated to be optimal, wherein the model is updated as shown in the formula (9):
wherein theta represents a trainable model parameter, t-1 represents a previous time step, eta is a learning rate,
and
the corrected first moment deviation and the corrected second moment deviation are respectively obtained, and epsilon is a constant. The context-aware service demand dynamic prediction algorithm is shown as algorithm 1. When the AMEDIN model carries out service requirement prediction of context awareness, the input data of the AMEDIN model is set as
Wherein, UFoA feature that is representative of the current user,
representing a plurality of scene features, wherein only one scene feature in the m scene features is the scene feature where the user is located currently, and the rest scene features are empty; SRFtIs the t-th service requirement characteristic.
During data processing, scenes are already processed, 50 scenes used by the same user when the same service is used are taken, and less than 50 scenes are filled with 0 in the back. The 50 scenes of the user are not time series, so that which position the current scene is placed has no influence on the prediction result.
Further, in this embodiment, experiments and analysis of experimental results are performed on the service demand dynamic prediction method based on context awareness, which are specifically as follows:
(1) experimental data preparation
Currently, there is no data set for validating context-aware dynamic predictions of service demand. In order to verify the effectiveness of the method provided in this embodiment, a MovieLens dataset and an Alibaba dataset provided by a sky pool website are used in this embodiment. Wherein the Movielens data contains 100 ten thousand sample data from 6000 users scoring 18 categories of 4000 movies; the Alibaba dataset is data for a large number of users' advertisement click-through rates provided by the company arizaba. Regarding the MovieLens data, regarding the user characteristics as the user characteristics in service demand prediction, mapping the scene characteristics as the scene characteristics in service demand prediction, mapping the movie characteristics as the service demand characteristics in service demand prediction, and mapping the evaluation value of the movie of the user as a label of the user service demand; regarding the Alibaba data, the user characteristics are regarded as the user characteristics in service demand prediction, the scene characteristics are mapped to the scene characteristics in service demand prediction, the advertisement characteristics are mapped to the service demand characteristics in service demand prediction, and the value of whether the user clicks the advertisement is mapped to the label in user service demand prediction. In the experiment, the scoring category in the Movielens data is converted into two categories in the present embodiment, the original user score of the movie is a continuous value from 1 to 5, the samples with the scores of 4 and 5 are marked as positive in the present embodiment, and the rest are marked as negative. Since the data size of Alibaba is relatively large, 100 ten thousand pieces of data are randomly sampled from the data to serve as experimental data in the embodiment. In the experiment, 5 scene features in each data set are considered with emphasis, information in the Movielens data is shown in table 1, and information in the Alibaba data is shown in table 2.
Table 1 information in Movielens data
TABLE 2 information in Alibaba data
The experimental environment is a personal computer, the operating system is 64 bits of Windows 10 professional edition, the CPU is Intel i78750H, and the RAM is 8 GB. In the experiment, an open-source TensorFlow 2.0 GPU is selected as an implementation framework of a prediction model, and a Python 3.6 programming is adopted to implement the prediction model.
(2) Evaluation index
In this embodiment, the deviation between the predicted value and the true value of the service requirement is calculated by using the root mean square error RMSE, the average absolute error MAE, and the accuracy Acc, so as to measure the effectiveness of the prediction method provided in this embodiment. The formula for calculating RMSE and MAE is shown in formula (10) and formula (11).
Wherein, the smaller the values of RMSE and MAE are, the higher the prediction precision of the model is; a larger value of Acc indicates a higher prediction accuracy of the model.
(3) AMEDIN model parameter setting
In the AMEDIN model, an interaction unit is used for learning interaction relations between a plurality of scenes and service requirements, the number of network layers of the interaction unit has important influence on the performance of the model, and in order to enable the AMEDIN model to have better performance, the optimal value of the number of layers is determined through an experimental method. In the experiment, Adam is adopted as an optimization algorithm, and the initial learning rate is 0.00001. And observing the performance of the AMEDIN model by setting different layer numbers, and further determining the optimal value of the network layer number in the interaction unit. The results of the experiment are shown in table 3.
TABLE 3 Effect of the number of layers of the Interactive element network on the AMEDIN model Performance
As can be seen from table 3, increasing the number of layers of the interactive units helps to improve the model prediction accuracy. However, as the number of layers increases, the degree of improvement in model performance is limited. Meanwhile, the increase of the number of layers of the interaction units brings more parameter learning overhead to the model, so that the complexity of model training and the overfitting risk are increased. Based on the above experimental results, the performance and training consumption of the model are considered comprehensively, and in this embodiment, the number of network layers of the interaction unit is determined to be 4.
In addition, in the user service demand prediction model, the number of neuron nodes in each layer of the interaction unit has a large influence on the performance of the model, and in order to enable the model to have a good prediction capability, in the amedn model, the number of neuron nodes in each layer of the network is respectively set to be 16, 32, 64, 128 and 256, the model is executed, and the experimental result is recorded, and is shown in table 4.
TABLE 4 influence of interaction unit node number on AMEDIN model Performance
As can be seen from table 4, as the number of neuron nodes increases, the performance of the model is gradually improved; when the number of the nodes reaches 128, the performance of the model reaches the optimum; thereafter, as the number of nodes increases, the performance of the model all degrades. Meanwhile, increasing the number of nodes of the neuron increases the overhead of model training and the risk of overfitting. Based on the experimental result, in this embodiment, the number of nodes of each layer of network neurons in the amedn model interaction unit is set to 128.
In the user service demand prediction model, parameters of the model are optimized through an Adam algorithm, wherein the learning rate of the Adam algorithm has a large influence on the stability of the prediction model, in order to enable the model to have good prediction capability, the learning rate of the experiment is respectively set to be 1e-2, 1e-3, 1e-4 and 1e-5 in the AMEDIN model, the model is executed, and the experiment result is recorded, wherein the experiment result is shown in fig. 2 and fig. 3.
As can be seen from fig. 2 and 3, as the learning rate pair decreases, the behavior of the model gradually stabilizes; as can be seen from FIG. 2, when the learning rate of the AMEDIN model on the Movielens data set is greater than 1e-5, the model is overfitting, and the model parameters cannot be optimized. As can be seen from fig. 3, although the amedan model does not have a serious overfitting phenomenon on the Alibaba data set, the difference between the verification loss of the model and the training loss is large, and the optimal model parameters cannot be learned by better fitting the data distribution. Based on the experimental results, the learning rate of the amedn model is set to 1e-5 in the present embodiment.
(4) Setting of contrast model parameters
To verify the effectiveness of the proposed method, five typical deep neural networks were selected for comparison with the amedn model constructed in this example. Five typical deep neural networks are: deep Neural Networks (DNNs), deep fms, Attention Interaction Networks (AIN), Attention Factorization Machines (AFMs), and Nerve Factorization Machines (NFMs).
For the deep neural network DNN, user characteristics, scene characteristics and service requirement characteristics are input into a full-connection network after passing through an Embedding layer (Embedding) respectively, and the service requirement of a user is predicted. In the experiment, the number of hidden layers is set to be 3, the number of nodes of the hidden layers is set to be 32, a cross entropy loss function is adopted, an Adam optimization algorithm is adopted, and the initial learning rate is set to be 0.001. For deep FM, after the user characteristics, the scene characteristics, and the service requirement characteristics are processed by an embedding layer, a Factorization Machine (FM) is used to extract low-order characteristics, and then DNN is used to perform high-order characteristic extraction, and then the extracted low-order characteristics are input to a full-connection network to predict the service requirement of the user. In the experiment, parameter setting of deep fm is consistent with that of the original paper, Dropout is set to be 0.5, Adam is selected as an optimization algorithm, and the initial learning rate is set to be 0.001. For AIN, the number of hidden layers is set to be 2, the number of nodes of the hidden layers is set to be 128, Adam is adopted as an optimization algorithm, and the initial learning rate is set to be 0.001.
For the AFM model, Dropout is set to be 0.5 according to the original paper, a batch training strategy with the size of 512 is used, Adam is selected as an optimization algorithm, and the initial value of the learning rate is set to be 0.001. The NFM is a neural network model for sparse data prediction, a factorization machine is enhanced under the neural network model to learn high-order interactive features, a batch training strategy with the size of 512 is set, Adam is adopted as an optimization algorithm, and the initial learning rate is 0.01. For the amedn model proposed in this embodiment, the number of layers of the interactive unit is set to 4, the number of nodes of the hidden layer is set to 128, Adam is used as an optimization algorithm, and the initial learning rate is 0.00001. For each network model, an Alibaba data set and a Movielens data set are used as experimental data, and the maximum iteration number is set to be 300.
(5) Comparison of Performance of different models
In order to verify the effectiveness of the method proposed in this embodiment, the experiment uses 80% of the data set as training data and 20% of the data set as test data, and multiple models are trained and tested. The evaluation index given above is used to measure the performance of each model. The results of the experiment are shown in Table 5.
TABLE 5 Performance evaluation of different models on two sets of data sets
As can be seen from table 5, when service demand prediction is performed, the amedn model provided in this embodiment is superior to other methods in the evaluation indexes Acc, RMSE, and MAE. In the Movielens data set, the AMEDIN model is respectively superior to the optimal result 1 in other methods in the evaluation index Acc. 14 percent; on the indices RMSE and MAE, 0.51% and 0.5% lead the suboptimal results, respectively. In the Alibaba data set, the AMEDIN model is respectively superior to the optimal result 1 in other methods on the evaluation index Acc. 48%, leading the suboptimal result 1 on the indicators RMSE and MAE, respectively. 06% and 1. 6 percent. According to the experimental results, the AMEDIN model provided by the embodiment can effectively capture different influences of a plurality of scenes on service requirements through the interaction unit and the attention mechanism; meanwhile, by extracting the interactive characteristics of a plurality of scenes and service requirements, the loss of the nonlinear relation between the scenes and the service requirements can be effectively reduced; on the other hand, the scene characteristics with the largest influence weight on the user service demand are obtained through the attention mechanism, and the prediction precision of the model is improved.
(6) Convergence verification of different models
In order to verify the convergence of the method proposed in this embodiment, a plurality of models are trained and tested, respectively, and the parameter settings of the plurality of models are consistent with the above settings. The results of the experiment are shown in fig. 4 and 5. Wherein the ordinate represents the prediction accuracy and the abscissa represents the number of iterations of the model.
As can be seen from fig. 4 and 5, as the number of iterations increases, the prediction accuracy of the plurality of deep neural networks is continuously improved. As can be seen from fig. 4, the NFM model performed the weakest in the Movielens dataset. As can be seen from fig. 5, the performance of the DNN model is the weakest in the Alibaba dataset. The AMEDIN model provided in the embodiment has better prediction accuracy on two data sets. The AMEDIN model has good learning ability, and higher prediction accuracy can be obtained with fewer iterations.
(7) Analysis of influence of different modules on AMEDIN model performance
In order to verify the effectiveness of obtaining the interaction relationship and obtaining the influence weight for improving the service demand prediction accuracy, the experiment obtains different variants of the AMEDIN model by removing related operations, and verifies the influence of different modules on the performance of the AMEDIN model by comparing the variants of the AMEDIN model with the AMEDIN model. Wherein, the AMEDINNoIta representation does not consider the interaction relation between different scenes and service requirements, the AMEDINNoPooling representation does not consider the interaction relation between the pooled scene features and the service requirements, and the AMEDINNoAtt representation does not use an attention mechanism. The AMEDIN model is consistent with the parameter settings of other variant models, the four models are trained and tested based on the same data set and an experimental platform, and the experimental results are shown in fig. 6 to 9, wherein the ordinate represents RMSE and MAE respectively, and the abscissa represents the iteration times of the algorithm.
As can be seen from fig. 6 to 9, when the amedn model does not consider the interaction relationship and the amedn model does not use the attention mechanism, the variant model of amedn has an increased value of the RMSE and MAE evaluation indexes, and the amedn model is superior to the other three variant models in the RMSE and MAE indexes. Based on the experimental results, it can be obtained that the interaction relationship between different scenes and service requirements and the influence weight of the different scenes on the service requirements are considered, which is helpful for improving the performance of the AMEDIN model, that is, explaining, the interaction relationship between the scenes and the service requirements and the influence weight of the different scenes on the service requirements are considered, and the accuracy of service requirement prediction is facilitated to be improved.
In order to realize context-aware dynamic prediction of service requirements and further improve the initiative and intelligence of service recommendation, the embodiment provides a context-aware dynamic prediction method of service requirements. The research work constructs an attention mechanism enhanced deep interaction neural network model AMEDIN, and firstly, the interaction relation between a plurality of different scenes and service requirements is captured through an interaction unit in the AMEDIN model; then, acquiring the influence weight of a plurality of scenes on the service demand through an attention mechanism; and finally, constructing training data based on the output of the attention mechanism, and dynamically predicting the service requirement through a full-connection network. A large number of experiments are carried out based on a real data set, and the effectiveness of the method provided by the embodiment is verified; meanwhile, the effectiveness of the interaction unit and the attention mechanism on improving the model prediction accuracy is verified through experiments. In subsequent research work, main factors influencing the service demand prediction precision are further analyzed, influence rules of the factors on the service demand prediction are mined, and research on the service demand prediction is carried out based on the rules, so that the flexibility and the accuracy of the service demand prediction are improved.
Example two:
the embodiment aims to provide a dynamic service demand prediction system based on context awareness
A dynamic prediction system of service demand based on context awareness, comprising:
the data acquisition unit is configured to acquire relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information;
a service demand prediction unit configured to perform dynamic prediction of a service demand using a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
Example three:
the embodiment aims at providing an electronic device.
An electronic device comprising a memory, a processor and a computer program stored in the memory for execution by the processor, the processor implementing a context awareness based dynamic prediction system of service demand when executing the program, comprising:
acquiring relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information; dynamically predicting the service requirement by utilizing a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
Example four:
it is an object of the present embodiments to provide a non-transitory computer-readable storage medium.
A non-transitory computer readable storage medium having stored thereon a computer program which when executed by a processor implements a context awareness based service demand dynamic prediction system, comprising:
acquiring relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information; dynamically predicting the service requirement by utilizing a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
Although the present disclosure has been described with reference to specific embodiments, it should be understood that the scope of the present disclosure is not limited thereto, and those skilled in the art will appreciate that various modifications and changes can be made without departing from the spirit and scope of the present disclosure.
Claims (10)
1. A service demand prediction method based on context awareness is characterized by comprising the following steps:
acquiring relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information; dynamically predicting the service requirement by utilizing a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
2. The method as claimed in claim 1, wherein for obtaining the interaction relationship between different scenarios and service requirements, the data needs to be preprocessed before being input into the interaction unit, and first, for each user, the data with the same service requirements are grouped into one group; and then, extracting different scene characteristics in the same group of data, and combining the scene characteristics with the user characteristics and the service requirement characteristics to form a piece of sample data.
3. The method according to claim 1, wherein after the interactive relationship between different scene features and service requirement features is obtained, the scene features are pooled in an average pooling manner to obtain pooled scene features of a plurality of scene features, and the pooled scene features represent main scene features of a plurality of scenes initiating a service requirement; and acquiring an interactive relation between the pooling scene characteristics and the service requirement characteristics through the interactive unit.
4. The method as claimed in claim 1, wherein when calculating the influence weights of different scenes on the service demand, the scene features and the interaction relationships corresponding to the scene features and the pooling features are combined into a new stitching vector, the output vector through the attention mechanism is a weighted sum of the stitching vector and the influence weights, and the vector represents the scene features with obvious influence on the service demand initiated by the user.
5. The method as claimed in claim 4, wherein after the scene features having a large influence on the user initiated service demand are obtained, the prediction module based on the attention mechanism enhanced deep interaction neural network model realizes the prediction of the service demand.
6. The method of claim 1, wherein the attention mechanism-enhanced deep interaction neural network model selects a cross entropy loss function to optimize the constructed model.
7. The method as claimed in claim 1, wherein the interactive unit and the service demand prediction module are formed by a fully connected network.
8. A dynamic service demand prediction system based on context awareness, comprising:
the data acquisition unit is configured to acquire relevant data when a user puts forward a service demand, wherein the relevant data comprises characteristic information of the user, scene information of the user and the service demand information;
a service demand prediction unit configured to perform dynamic prediction of a service demand using a pre-trained attention mechanism enhanced deep interaction neural network model;
the network model comprises an interaction unit, an influence weight learning module and a service demand prediction module, and the interaction relation between different scenes and service demands is captured through the interaction unit; further learning influence weights of different scenes on service requirements based on an attention mechanism through an influence weight learning module; and finally, realizing service demand prediction according to the influence weight through a service demand prediction module.
9. An electronic device comprising a memory, a processor and a computer program stored in the memory for execution, wherein the processor implements a method for dynamic prediction of service demand based on context awareness as claimed in any one of claims 1 to 7.
10. A non-transitory computer readable storage medium, on which a computer program is stored, wherein the program, when executed by a processor, implements a method for dynamic prediction of service demand based on context awareness as claimed in any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011526415.4A CN112529637B (en) | 2020-12-22 | 2020-12-22 | Service demand dynamic prediction method and system based on context awareness |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011526415.4A CN112529637B (en) | 2020-12-22 | 2020-12-22 | Service demand dynamic prediction method and system based on context awareness |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112529637A true CN112529637A (en) | 2021-03-19 |
| CN112529637B CN112529637B (en) | 2022-12-06 |
Family
ID=75002265
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011526415.4A Active CN112529637B (en) | 2020-12-22 | 2020-12-22 | Service demand dynamic prediction method and system based on context awareness |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112529637B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113537623A (en) * | 2021-07-30 | 2021-10-22 | 烟台大学 | Attention mechanism and multi-mode based dynamic service demand prediction method and system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
| CN111079674A (en) * | 2019-12-22 | 2020-04-28 | 东北师范大学 | Target detection method based on global and local information fusion |
| CN111797321A (en) * | 2020-07-07 | 2020-10-20 | 山东大学 | Personalized knowledge recommendation method and system for different scenes |
-
2020
- 2020-12-22 CN CN202011526415.4A patent/CN112529637B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110879864A (en) * | 2019-10-29 | 2020-03-13 | 南京大学 | Context recommendation method based on graph neural network and attention mechanism |
| CN111079674A (en) * | 2019-12-22 | 2020-04-28 | 东北师范大学 | Target detection method based on global and local information fusion |
| CN111797321A (en) * | 2020-07-07 | 2020-10-20 | 山东大学 | Personalized knowledge recommendation method and system for different scenes |
Non-Patent Citations (2)
| Title |
|---|
| MEI L 等: "An Attentive Interaction Network for Context-aware Recommendations", 《CONFERENCE ON INFORMATION AND KNOWLEDGE MANAGEMENT》 * |
| 刘志中 等: "情景感知的物联网服务推荐方法研究", 《计算机应用研究》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113537623A (en) * | 2021-07-30 | 2021-10-22 | 烟台大学 | Attention mechanism and multi-mode based dynamic service demand prediction method and system |
| CN113537623B (en) * | 2021-07-30 | 2023-08-18 | 烟台大学 | Attention mechanism and multi-mode based service demand dynamic prediction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112529637B (en) | 2022-12-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111177575B (en) | Content recommendation method and device, electronic equipment and storage medium | |
| CN111797321B (en) | Personalized knowledge recommendation method and system for different scenes | |
| WO2021155706A1 (en) | Method and device for training business prediction model by using unbalanced positive and negative samples | |
| CN110751318B (en) | Ultra-short-term power load prediction method based on IPSO-LSTM | |
| CN112529638B (en) | Service demand dynamic prediction method and system based on user classification and deep learning | |
| CN108549658A (en) | A kind of deep learning video answering method and system based on the upper attention mechanism of syntactic analysis tree | |
| CN107451596B (en) | Network node classification method and device | |
| CN114780831A (en) | Sequence recommendation method and system based on Transformer | |
| CN113128671B (en) | Service demand dynamic prediction method and system based on multi-mode machine learning | |
| CN113628059A (en) | Associated user identification method and device based on multilayer graph attention network | |
| CN112256866A (en) | Text fine-grained emotion analysis method based on deep learning | |
| CN113706151A (en) | Data processing method and device, computer equipment and storage medium | |
| CN113837308A (en) | Knowledge distillation-based model training method and device and electronic equipment | |
| CN114490065A (en) | Load prediction method, device and equipment | |
| CN115695025B (en) | Training method and device for network security situation prediction model | |
| CN114781503A (en) | Click rate estimation method based on depth feature fusion | |
| CN112529637B (en) | Service demand dynamic prediction method and system based on context awareness | |
| CN108491477B (en) | Neural network recommendation method based on multi-dimensional cloud and user dynamic interest | |
| CN113935413A (en) | Distribution network wave recording file waveform identification method based on convolutional neural network | |
| CN113723108A (en) | Event extraction method and device, electronic equipment and storage medium | |
| CN116452241B (en) | User loss probability calculation method based on multi-mode fusion neural network | |
| CN112232079A (en) | Microblog comment data classification method and system | |
| CN116543289A (en) | Image description method based on encoder-decoder and Bi-LSTM attention model | |
| CN113705197B (en) | Fine granularity emotion analysis method based on position enhancement | |
| CN114357284A (en) | Crowdsourcing task personalized recommendation method and system based on deep learning |
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 |