CN112927037B - Vendor recommendation method and system - Google Patents
Vendor recommendation method and system Download PDFInfo
- Publication number
- CN112927037B CN112927037B CN202110156145.0A CN202110156145A CN112927037B CN 112927037 B CN112927037 B CN 112927037B CN 202110156145 A CN202110156145 A CN 202110156145A CN 112927037 B CN112927037 B CN 112927037B
- Authority
- CN
- China
- Prior art keywords
- recommendation
- function
- nonlinear
- user
- batch
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 30
- 230000005012 migration Effects 0.000 claims abstract description 22
- 238000013508 migration Methods 0.000 claims abstract description 22
- 238000013526 transfer learning Methods 0.000 claims abstract description 19
- 230000000694 effects Effects 0.000 claims abstract description 17
- 230000008447 perception Effects 0.000 claims abstract description 15
- 230000003044 adaptive effect Effects 0.000 claims abstract description 12
- 239000013598 vector Substances 0.000 claims abstract description 12
- 238000009826 distribution Methods 0.000 claims abstract description 3
- 230000006870 function Effects 0.000 claims description 43
- 238000013507 mapping Methods 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 238000012545 processing Methods 0.000 abstract description 2
- 230000008859 change Effects 0.000 description 18
- 238000009856 non-ferrous metallurgy Methods 0.000 description 6
- 230000008569 process Effects 0.000 description 6
- HCHKCACWOHOZIP-UHFFFAOYSA-N Zinc Chemical compound [Zn] HCHKCACWOHOZIP-UHFFFAOYSA-N 0.000 description 5
- 238000004519 manufacturing process Methods 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 239000011701 zinc Substances 0.000 description 5
- 229910052725 zinc Inorganic materials 0.000 description 5
- 230000008901 benefit Effects 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- XUIMIQQOPSSXEZ-UHFFFAOYSA-N Silicon Chemical compound [Si] XUIMIQQOPSSXEZ-UHFFFAOYSA-N 0.000 description 2
- 101000609272 Thermobifida fusca (strain YX) Phenylacetone monooxygenase Proteins 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 239000002994 raw material Substances 0.000 description 2
- 229910052710 silicon Inorganic materials 0.000 description 2
- 239000010703 silicon Substances 0.000 description 2
- 238000003723 Smelting Methods 0.000 description 1
- 102100029469 WD repeat and HMG-box DNA-binding protein 1 Human genes 0.000 description 1
- 101710097421 WD repeat and HMG-box DNA-binding protein 1 Proteins 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000010924 continuous production Methods 0.000 description 1
- 230000000593 degrading effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000011084 recovery Methods 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
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
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0605—Supply or demand aggregation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
The invention relates to the technical field of computers, and discloses a provider recommendation method and a provider recommendation system, which are used for processing nonlinear recommendation problems along with interest and feature changes of users and ensuring recommendation effects. The method comprises the following steps: receiving at least one vendor sample; predicting recommended operations corresponding to each supplier sample based on an adaptive multi-layer perceived online transfer learning model; tracking and acquiring actual adoption effects of the provider users on each provider sample; according to comparison of recommended operation and actual adoption effect, weight distribution and nonlinear parameter vectors between the user characteristics of the previous batch and the user characteristics of the current batch of the purchasing users served by the self-adaptive multi-layer perception online transfer learning model are adjusted; and re-acquiring new supplier samples, predicting corresponding recommended operations by using the adjusted online migration learning model, and repeating the steps until all the supplier samples meeting the supplier users are screened out.
Description
Technical Field
The present invention relates to the field of computer technologies, and in particular, to a vendor recommendation method and system.
Background
The nonferrous metallurgy provider selection problem is a multi-attribute decision-making problem, and a purchasing department of an enterprise selects a proper provider for purchasing according to the requirements of the enterprise, so that the benefit and efficiency of the enterprise are maximized. With the development of enterprises, the number of alternative suppliers for each purchasing activity is increased, and new suppliers are added. The characteristics of nonferrous metallurgy suppliers also become complex along with the increase of production requirements, and the decision-making efficiency of purchasing departments tends to be reduced for a large number of suppliers with complex information. Such problems can be attributed to information overload problems for the suppliers. On the other hand, the demand of the purchasing department for suppliers may change during the operation of nonferrous metallurgical enterprises. For enterprises requiring continuous production such as nonferrous metallurgy, the characteristics of suppliers determine the emphasis of enterprise production over a period of time. During this period, the enterprise is affected by factors such as working conditions and environment, and the demands of the enterprise on the suppliers are different at different times, and the demand change is rapid and unavoidable. For example, nonferrous metallurgical enterprises have a wide variety of requirements for raw material suppliers, and in the case of stable production and stock, the purchasing department will be more biased towards price factors. When the production condition is unstable, the purchasing department can prefer to supply the suppliers with stable quality. In this case, information overload of the suppliers will affect decision efficiency of the purchasing department more, and demand change of the purchasing department will complicate recommendation of the suppliers more. In the face of such a lack of knowledge of the purchasing department, the recommendation system is able to learn the needs of the purchasing department and generate appropriate supplier recommendations. Therefore, the construction of the recommendation system which can adapt to the demand change of the purchasing department and can process the information overload of the suppliers has very important practical significance for purchasing decisions of nonferrous metallurgy enterprises.
In the face of a large amount of project information, users need a recommendation system to improve the decision making efficiency. FIG. 1 illustrates a training process for a recommendation system with varying user needs. S represents the demand characteristics of the user, and x and y represent the recommended items of a single batch and the operation results (labels) of the user respectively. The recommendation system trains the user demand characteristics according to the operation result of each batch purchasing department, and gradually obtains stable demand characteristics. The recommendation system then predicts the new item based on the trained demand characteristics to complete the recommendation. However, during use, the user's needs may change and the recommendation system needs to be able to follow the user's needs to maintain the accuracy of the recommendation, which can be attributed to a concept drift problem. First, the user' S demand remains stable for an initial period of time, which means that the demand characteristics S of the training 1 =S 2 =…=S t+1 =s. At this time for any itemOrder x t Arbitrary S t The predicted outcome y of (2) t Are substantially identical. Next at a certain moment j, the user 'S operation changes such that the trained user' S demand S changes, S i ≠S j 。
However, the prior art has the following drawbacks:
1. nonferrous metallurgical suppliers recommend a nonlinear classification problem, which means that it must be solved by nonlinear methods.
2. Because the demand of the purchasing department is influenced by the outside in the running process, the demand changes. The method is to use online learning to follow the change of the user characteristics, and because the online learning only adjusts the user demand characteristics according to the current sample, the sample conflict problem caused by the old data set can be ignored in the instantaneity. However, due to an erroneous sample generated by a possible misoperation of the user, the user characteristics of online learning may fluctuate, thereby causing fluctuations in the recommendation accuracy.
Disclosure of Invention
The invention aims to disclose a provider recommendation method and a provider recommendation system, which are used for processing nonlinear recommendation problems along with interest and feature changes of users and ensuring recommendation effects.
To achieve the above object, the present invention discloses a vendor recommendation method, comprising:
receiving at least one vendor sample;
predicting recommended operations corresponding to each supplier sample based on an adaptive multi-layer perceived online transfer learning model;
tracking and acquiring actual adoption effects of the provider users on each provider sample;
according to comparison of recommended operation and actual adoption effect, weight distribution and nonlinear parameter vectors between the user characteristics of the previous batch and the user characteristics of the current batch of the purchasing users served by the self-adaptive multi-layer perception online transfer learning model are adjusted;
and re-acquiring new supplier samples, predicting corresponding recommended operations by using the adjusted online migration learning model, and repeating the steps until all the supplier samples meeting the supplier users are screened out.
Preferably, the prediction function of the online migration learning model based on the adaptive multi-layer perception is:
wherein y' t For the result of whether the operation of the recommendation system is recommended,as a projection function, alpha 1,t And alpha 2,t The weight parameters corresponding to the user characteristic weights of the previous batch and the user characteristic weights of the current batch are respectively,and->Is a nonlinear projection function based on multi-layer perception, l α Is a function of the single lot prediction error rate, Θ (v, v Φ ) For the end user characteristics of the previous batch, +.>Is a user characteristic of the current time of the batch.
Preferably, alpha 1,t And alpha 2,t Initial values of (2) are 0.5, respectively; the update function is:
wherein,u φ as a function of the non-linear mapping,η is the transfer learning rate; according to the success or failure of the recommendation result, the linear part of the current feature weight of the user is updated according to the following rule:
wherein,for the user feature updated at the next time, y t For the actual user operation corresponding to the actual adoption effect, < ->For the hinge loss function->For learning rate, β is w t Is used for updating the parameters of the speed.
Preferably, in the online migration learning model based on adaptive multi-layer perception, z is used t As hidden layer nodes of the nonlinear multi-layer aware network, from z t To z t+1 The hidden layer weight is updated in a changing mode of (a):
wherein,is the learning rate.
Preferably, in the online migration learning model based on adaptive multi-layer perception, the method for dividing the nonlinear mapping function into at least two linear components so as to enable the model to continue to use a parameter updating mode of online migration learning comprises the following steps:
the following nonlinear mapping function based on MLP is adopted:
z t =[z t 1 ,z t 2 ,z t 3 ,...,z t h ]
wherein z is t i Is the ith hidden layer node, and each node is in accordance with the ReLU functionAs an activation function, h is the number of hidden layer nodes; the parameter vectors defining the nonlinear mapping function are:
φ t =[r 1 t ,r 2 t ,...,r h t ]
the update mode of the parameter vector obeys the update strategy of the PA regression algorithm as follows:
wherein the method comprises the steps of,Loss function for PA regression algorithm, +.>To learn the rate, r i For every i elements of hidden layer weights.
Preferably: l (L) α =e f(MC) ,
Wherein MC is the single-batch prediction error rate,for preference parameters, max (MC) is the current maximum prediction error rate, and Min (MC) is the current minimum prediction error rate.
The invention also discloses a supplier recommending system, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor realizes the corresponding steps of the method when executing the computer program.
The invention has the following beneficial effects:
1. more complex nonlinear recommendation problems can be handled.
2. The method has strong robustness to misoperation of the purchasing department, and ensures that the recommendation accuracy is kept stable under the condition of unchanged requirements.
3. Under the condition of changing the demand, the online migration learning model based on the self-adaptive multi-layer perception can more quickly follow the interest change of the user, so that the long-term recommendation effect is improved.
The invention will be described in further detail with reference to the accompanying drawings.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention. In the drawings:
FIG. 1 is a training process of a prior art recommendation system with varying user needs.
Fig. 2 is a schematic diagram of a recommendation system architecture based on multi-layer perceptron (MLP) adaptive online migration learning according to an embodiment of the present invention.
Fig. 3 is a schematic diagram showing the comparison of recommendation accuracy of the recommendation system (AOTLMLP) and other recommendation systems according to the present invention.
Detailed Description
Embodiments of the invention are described in detail below with reference to the attached drawings, but the invention can be implemented in a number of different ways, which are defined and covered by the claims.
Example 1
The embodiment aims to solve the problem of long-term recommendation of the raw material suppliers of nonferrous metallurgy enterprises, so that the built recommendation system has a better recommendation effect. For nonferrous metallurgy purchasing departments, firstly, the feature classification of suppliers is nonlinear; secondly, the demands of enterprises on ore suppliers are different in different time periods, and decision departments can be influenced by the forecast values of the suppliers to make decisions (misoperation) which do not meet the demands in the purchasing process. In the embodiment, a recommendation system is built by selecting a learning algorithm of online migration, and nonlinear improvement and self-adaptive improvement based on multilayer perception are made on the basis of the recommendation system. Making the recommendation system more suitable for recommendation by nonferrous metallurgical suppliers.
The structure of the recommendation system based on multi-layer perceptron (MLP) self-adaptive online migration learning in this embodiment is shown in fig. 2, and AOTLMO is a recommendation algorithm module. The user profile for the corresponding buyer in the AOTLMO contains two components, linear and nonlinear. Wherein Θ (v, v Φ ) For the previous batch of end user features, the portion of the fixed features do not change with online learning.For the user characteristic of the current time of the batch, +.>User characteristics updated for the next time; the part before comma in the user characteristic button is a linear characteristic, and the part after comma is a nonlinear characteristic. In content-based recommendations, the vendor features of the same batch may be represented by a feature vector x= [ X ] 1 x 2 x 3 … x n ]To represent. Wherein x is t Features representing the t-th vendor. In the feature vector, each feature is quantized, and for those tag features that are not quantized, we can represent the membership of the corresponding item to that feature; the corresponding operation is a digitalized pretreatment of the supplier features, which is common knowledge of the person skilled in the art and will not be described in detail. The output of the recommendation system is denoted by Y, Y ε { 1,1},1 representing a recommendation, and-1 representing a non-recommendation.
The nonlinear self-adaptive online transfer learning recommendation algorithm can adjust the feature weight according to the actual operation of each batch of users and the prediction result of the recommendation algorithm. First, let v and w t The linear portions of the user characteristics of the previous batch and the user characteristics of the current t batch, respectively. Then given an item x t The recommendation system may predict its tag as:
wherein y' t For the result of whether the operation of the recommendation system is recommended,is a projection function. Alpha 1,t And alpha 2,t The initial values of the weight parameters corresponding to the historical feature weight and the current feature weight are 0.5 respectively.And->Is a nonlinear projection function based on multi-layer perception. With continuous input of samples,α 1,t And alpha 2,t Is dynamically changing, its update function is as follows:
wherein,u φ is a nonlinear mapping function. But->η is the transfer learning rate. According to the success or failure of the recommendation result, the linear part of the current feature weight of the user is updated according to the following rule:
wherein y is t For the actual user operation to be performed,for the hinge loss function->Is the learning rate. Again, we introduce a beta parameter to limit w t Is directed to roughly classifying the nonlinear dataset. After rough classification, the rest of the classification task is completed by the nonlinear section. We use z t De-represent and use it as a hidden layer node of the nonlinear multi-layer aware network. Then we can get z from the PA algorithm t Is updated by the update method of (a).
Wherein,is the learning rate. By the above steps we obtain the product from z t To z t+1 This can be used to solve the problem of updating hidden layer weights.
Further, considering the necessity of online learning, we split the nonlinear mapping function into multiple linear components so that the algorithm can continue to use the parameter update approach of online transfer learning. We therefore propose the following non-linear mapping function based on MLP:
z t =[z t 1 ,z t 2 ,z t 3 ,...,z t h ]
wherein z is t i Is the ith hidden layer node, and each node is provided with(ReLU) as an activation function, and h is the number of hidden nodes. The ReLU function as an activation function of the multi-layer aware network may represent an arbitrary function in case of sufficient number of layers and nodes. And, the ReLU function can be divided into two sections of linear functions, which allows hidden layer weights to be updated according to the original linear algorithm. We define the parameter vector of the nonlinear mapping function as:
φ t =[r 1 t ,r 2 t ,...,r h t ]
the update of the parameter vector is performed in accordance with the update strategy of the PA regression algorithm as follows, whereinLoss function for PA regression algorithm, +.>Is the learning rate.
In the above, r i For every i elements of the hidden layer weight, whereby each element of the hidden layer weight may be updated according to the above update policy.
The recommendation algorithm based on nonlinear online transfer learning comprehensively considers the historical characteristics and the current characteristics of the previous period. The historical characteristics determine the stability of the recommendation accuracy when the demand is stable, while the current characteristics determine the following speed of the recommendation system to the demand. In the case of constant user demand, the predictive function is biased towards historical features to make the synthesis more stable. However, in response to the user demand changing more rapidly, the recommendation system needs to make the recommendation accuracy after the demand change reach the standard at a faster speed to ensure the practicability of the system. On-line transfer learning, if focusing on the learning speed of the recommendation system when the demand changes, the robustness of the recommendation system to erroneous samples when the demand is stable is reduced. In order to enable the online migration recommendation algorithm to distinguish the demand change from the error sample and enable the system to achieve self-adaption, the embodiment provides the self-adaption online migration learning recommendation algorithm and further introduces the following self-adaption loss function.
In order to achieve both the transfer learning speed and the robustness to the erroneous sample, the present embodiment needs to enable the recommendation system to identify the difference between the erroneous sample and the sample that is changed when the real demand is changed. The main difference between the two is that the error samples are a small number of intermittent error classifications during the whole training process, and the sample changes when the demands change are a large number of continuous error classifications. To this end we introduce the following loss functions:
l α =e f(MC)
wherein MC is the single-batch prediction error rate,for preference parameters, max (MC) is the current maximum prediction error rate, min (MC) is the current minimum prediction error rate, and the modified prediction function is as follows:
l α is a function of the single lot prediction error rate, l in the case of small single lot error rate α The original online migration learning effect is not affected, and the stability of the recommendation accuracy in the state of unchanged requirements is maintained. Under the condition of large error rate of single batch, the algorithm considers that the change of the demand occurs, l α Will increase and thereby speed up the adjustment of the demand characteristics of the purchasing department. l (L) α The introduction of the adaptive loss function enables online migration learning to discern erroneous samples and differences in the training sequence for demand change.
In summary, the parameter adjustment of the present embodiment includes the following aspects:
learning recommendation algorithm based on self-adaption online migrationIn the method, parameters to be set include a loss function limiting parameter beta, a transfer learning rate eta and preference parametersFirst for the choice of β, a more complex nonlinear dataset needs to reduce β to allow more classification tasks to be done in the nonlinear section. In setting η, it is necessary to consider the item information complexity (the number of features, the number of data pieces). And when the demand is stable, the eta is regulated to ensure that the recommendation accuracy is maintained in a section so as to ensure the robustness of the recommendation system, and the recommendation accuracy can be recovered to a normal level after the demand is changed. Too small eta may ignore the impact of historical feature weights, degrading the recommendation system robustness, while too large eta may result in system divergence. />The stability and migration capabilities of the recommendation system are determined. If the stability requirements of the recommendation system on the recommendation accuracy are higher, then the +.>Is a value of (2). Conversely, if the recommender system emphasizes the following speed to the user's needs, then the +.>Is a value of (2). In classification questions->Values between 0 and 1 are possible.
Thus, the embodiment discloses a specific recommended algorithm step as follows:
input: all item feature sets { x ] of the batch 1 … x T -and previous stage end user features V, V Φ The previous lot prediction error rate MC, preference parameterLoss function limiting parameter beta, transfer learning rateη, batch size T.
Initializing: w (w) 1 =0,α 1,t =α 2,t =0.5,Φ t =[0.5,0.5,...,0.5]
Step 1: receiving a sample x t ;
Step 2: calculating z t And predicts the operation y 'corresponding to the sample' t ;
z t =[z t 1 ,z t 2 ,z t 3 ,...,z t h ]
Receiving the actual operation y of the sample by the user t ;
Number of mispredictions if classification is wrong: m=m+1; otherwise, returning to the second step;
step 3: calculating a new:
and to w t ,z t Updating:
step 4: according to z t Is a rate of change versus nonlinear parameter vector phi t Updating:
step 5: repeating the steps until T, and calculating the error prediction rate of the batch:
to verify the effectiveness of the adaptive online migration learning provider recommendation algorithm of this embodiment, we used purchase decisions and purchase data for a period of time in a zinc smelter purchasing department, and the feature set of the provider as a case. For the characteristics of the suppliers, such as economic factors, environmental factors, social factors, etc., the purchasing department will select a batch of suppliers that are currently suitable for the production needs for a period of time, thereby maximizing the enterprise benefit. For zinc smelting suppliers, 11 features are considered here, namely price level, zinc content, lead content, silicon content, zinc forecast value accuracy, lead forecast value accuracy, silicon forecast value accuracy, date of delivery accuracy, supply standard rate, natural geographic environment and enterprise development prospect. By tuning the parameters, we consider that, for a certain zinc smelter dataset, when β=0.9, η=0.05,the effect of the recommendation system is optimal. In this case we compare the recommendation effect of the recommendation system constructed separately from the original online migration algorithm and the nonlinear adaptive online migration algorithm and the PAMO nonlinear online algorithm, as shown in fig. 3.
Compared with the traditional recommendation system based on OTL, the recommendation system based on AOTLMO can handle the classification problem of the nonlinear data set, and therefore recommendation can be completed better. And compared with a recommendation system based on a common nonlinear online algorithm (PAMO), the AOTLMO recommendation system can relieve recommendation accuracy fluctuation caused by misoperation. Compared with a recommendation system formed by an online migration algorithm (OTLMO) based on MLP, the AOTLMO recommendation system can follow requirements more quickly after the requirements change and ensure quick recovery of recommendation accuracy. The recommended accuracy of the AOTLMO algorithm after the change of the demand can be recovered to more than 80% after about 20 batches, while the recommended accuracy of the OTLMO algorithm is always at a lower level, and the accuracy begins to recover after 50 batches. These results indicate that the AOTLMO recommendation system not only can keep stable and high recommendation accuracy under the condition of stable requirements, but also can adapt to user characteristics more quickly after the requirements change, and the system has the advantages of combining historical data and online adjustment capability.
Example 2
The embodiment discloses a provider recommendation system, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the computer program to realize the steps corresponding to the method in the embodiment.
In summary, the provider recommendation method and system disclosed by the embodiment of the invention have the following beneficial effects:
1. more complex nonlinear recommendation problems can be handled.
2. The method has strong robustness to misoperation of the purchasing department, and ensures that the recommendation accuracy is kept stable under the condition of unchanged requirements.
3. Under the condition of changing the demand, the online migration learning model based on the self-adaptive multi-layer perception can more quickly follow the interest change of the user, so that the long-term recommendation effect is improved.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (2)
1. A vendor recommendation method, comprising:
receiving at least one vendor sample;
predicting recommended operations corresponding to each supplier sample based on an adaptive multi-layer perceived online transfer learning model;
tracking and acquiring actual adoption effects of the provider users on each provider sample;
according to comparison of recommended operation and actual adoption effect, weight distribution and nonlinear parameter vectors between the user characteristics of the previous batch and the user characteristics of the current batch of the purchasing users served by the self-adaptive multi-layer perception online transfer learning model are adjusted;
re-acquiring new supplier samples, predicting corresponding recommended operations by using the adjusted online migration learning model, and repeating the steps until all the supplier samples meeting the supplier users are screened out;
the prediction function of the online transfer learning model based on the self-adaptive multi-layer perception is as follows:
wherein y' t For the result of whether the operation of the recommendation system is recommended,as a projection function, alpha 1,t And alpha 2,t The weight parameters corresponding to the user characteristic weights of the previous batch and the user characteristic weights of the current batch are respectively,and->Is a nonlinear projection function based on multi-layer perception, l α Is a function of the single lot prediction error rate, Θ (v, v Φ ) For the end user characteristics of the previous batch, +.>The method is characterized in that the method is a user characteristic of the current time of the batch, and is used for indicating that a part before comma in a user characteristic buckle number is a linear characteristic, and a part after comma is a nonlinear characteristic;
α 1,t and alpha 2,t Initial value divisionThe difference is 0.5; the update function is:
wherein,u∈R d ,u φ is a nonlinear mapping function +.>η is the transfer learning rate; according to the success or failure of the recommendation result, the linear part of the current feature weight of the user is updated according to the following rule:
wherein,for the user feature updated at the next time, y t For the actual user operation corresponding to the actual adoption effect, < ->For the hinge loss function->For learning rate, β is w t More of (2)Parameters of the new speed;
the vendor recommendation method further comprises:
in the online migration learning model based on adaptive multi-layer perception, z is used t As hidden layer nodes of the nonlinear multi-layer aware network, from z t To z t+1 The hidden layer weight is updated in a changing mode of (a):
wherein,is the learning rate;
the vendor recommendation method further comprises:
in the online transfer learning model based on adaptive multi-layer perception, the parameter updating mode for dividing the nonlinear mapping function into at least two linear components so that the model can continue to use online transfer learning comprises the following steps:
the following nonlinear mapping function based on MLP is adopted:
z t =[z t 1 ,z t 2 ,z t 3 ,...,z t h ]
wherein z is t i Is the ith hidden layer node, and each node is in accordance with the ReLU functionAs an activation function, h is the number of hidden layer nodes; definition of the definitionThe parameter vector of the nonlinear mapping function is:
φ t =[r 1 t ,r 2 t ,...,r h t ]
the update mode of the parameter vector obeys the update strategy of the PA regression algorithm as follows:
wherein,loss function for PA regression algorithm, +.>To learn the rate, r i Every i elements being hidden layer weights;
l α =e f(MC)
wherein MC is the single-batch prediction error rate,for preference parameters, max (MC) is the current maximum prediction error rate, and Min (MC) is the current minimum prediction error rate.
2. A vendor recommendation system comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the vendor recommendation method of claim 1 when executing the computer program.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110156145.0A CN112927037B (en) | 2021-02-04 | 2021-02-04 | Vendor recommendation method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110156145.0A CN112927037B (en) | 2021-02-04 | 2021-02-04 | Vendor recommendation method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112927037A CN112927037A (en) | 2021-06-08 |
CN112927037B true CN112927037B (en) | 2024-02-06 |
Family
ID=76170393
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110156145.0A Active CN112927037B (en) | 2021-02-04 | 2021-02-04 | Vendor recommendation method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112927037B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115358781B (en) * | 2022-08-22 | 2023-04-07 | 陕西师范大学 | Crowd sensing noise monitoring task recommendation method based on limited rational decision model |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016122A (en) * | 2017-04-26 | 2017-08-04 | 天津大学 | Knowledge recommendation method based on time-shift |
CN109190053A (en) * | 2018-07-04 | 2019-01-11 | 南京邮电大学 | One kind being based on point of interest importance and the authoritative point of interest recommended method of user |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8630960B2 (en) * | 2003-05-28 | 2014-01-14 | John Nicholas Gross | Method of testing online recommender system |
US10255628B2 (en) * | 2015-11-06 | 2019-04-09 | Adobe Inc. | Item recommendations via deep collaborative filtering |
-
2021
- 2021-02-04 CN CN202110156145.0A patent/CN112927037B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016122A (en) * | 2017-04-26 | 2017-08-04 | 天津大学 | Knowledge recommendation method based on time-shift |
CN109190053A (en) * | 2018-07-04 | 2019-01-11 | 南京邮电大学 | One kind being based on point of interest importance and the authoritative point of interest recommended method of user |
Non-Patent Citations (1)
Title |
---|
基于MLP 改进型深度神经网络学习资源推荐算法;樊海玮等;《计算机应用研究》;第37卷(第9期);第2629-2633页 * |
Also Published As
Publication number | Publication date |
---|---|
CN112927037A (en) | 2021-06-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Cheung et al. | Hedging the drift: Learning to optimize under nonstationarity | |
Torabi et al. | ABC inventory classification in the presence of both quantitative and qualitative criteria | |
Liao et al. | An integrated fuzzy TOPSIS and MCGP approach to supplier selection in supply chain management | |
Wang et al. | Developing a fuzzy TOPSIS approach based on subjective weights and objective weights | |
Soylu et al. | Multi-criteria inventory classification with reference items | |
Aguilar-Lasserre et al. | An AHP-based decision-making tool for the solution of multiproduct batch plant design problem under imprecise demand | |
Wang et al. | Schedule-based execution bottleneck identification in a job shop | |
CN112927037B (en) | Vendor recommendation method and system | |
US11755951B2 (en) | Machine learning with an intelligent continuous learning service in a big data environment | |
Demizu et al. | Inventory management of new products in retailers using model-based deep reinforcement learning | |
Arredondo et al. | Learning and adaptation of a policy for dynamic order acceptance in make-to-order manufacturing | |
US20190378061A1 (en) | System for modeling the performance of fulfilment machines | |
Rahman et al. | Supplier selection using fuzzy-topsis method: A case study in a cement industry | |
Karimi et al. | Towards optimal active learning for matrix factorization in recommender systems | |
Wang et al. | A neuro-fuzzy based forecasting approach for rush order control applications | |
Adam et al. | Machine learning for demand estimation in long tail markets | |
Veloso et al. | Self hyper-parameter tuning for stream recommendation algorithms | |
CN113744009A (en) | Target object output method and device, computer readable medium and electronic equipment | |
JPH10124478A (en) | Device for constructing segment generation type predicted model and method therefor | |
Höller et al. | Determining optimal parameters for expediting policies under service level constraints | |
TW202309810A (en) | Reducing sample selection bias in a machine learning-based recommender system | |
Liu | Using geometric programming to profit maximization with interval coefficients and quantity discount | |
Casanovas et al. | Decision making processes of non-life insurance pricing using fuzzy logic and OWA operators | |
Ali et al. | Personalized choice model for forecasting demand under pricing scenarios with observational data—The case of attended home delivery | |
JPWO2022070251A5 (en) |
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 |