WO2023050232A1 - 资产价值评估方法、模型训练方法、装置及可读存储介质 - Google Patents
资产价值评估方法、模型训练方法、装置及可读存储介质 Download PDFInfo
- Publication number
- WO2023050232A1 WO2023050232A1 PCT/CN2021/121958 CN2021121958W WO2023050232A1 WO 2023050232 A1 WO2023050232 A1 WO 2023050232A1 CN 2021121958 W CN2021121958 W CN 2021121958W WO 2023050232 A1 WO2023050232 A1 WO 2023050232A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- asset
- historical
- value
- user
- assets
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 126
- 238000012549 training Methods 0.000 title claims abstract description 123
- 238000011156 evaluation Methods 0.000 title claims abstract description 58
- 239000013598 vector Substances 0.000 claims abstract description 214
- 230000003993 interaction Effects 0.000 claims abstract description 84
- 230000008569 process Effects 0.000 claims description 58
- 230000006870 function Effects 0.000 claims description 22
- 238000004590 computer program Methods 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 238000010586 diagram Methods 0.000 description 12
- 238000010276 construction Methods 0.000 description 5
- 238000012986 modification Methods 0.000 description 5
- 230000004048 modification Effects 0.000 description 5
- 230000008520 organization Effects 0.000 description 5
- 238000012545 processing Methods 0.000 description 4
- 238000006243 chemical reaction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000002452 interceptive effect Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000001902 propagating effect Effects 0.000 description 2
- 238000012546 transfer Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000013210 evaluation model Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
Definitions
- the present disclosure relates to the field of computer technology, in particular to an asset value evaluation method, a model training method, a device, and a readable storage medium.
- the present disclosure provides an asset value evaluation method, a model training method, a device and a readable storage medium for improving the accuracy of asset value evaluation.
- an embodiment of the present disclosure provides an asset value evaluation method, which includes:
- the asset embedding vector of each of the assets is input into the graph convolutional network model to obtain the value of each of the assets to the user.
- the training process based on the relationship between each of the assets and attributes includes:
- the initializing the embedding vector of the triplet of the first type includes:
- the training process based on the relationship between each of the assets and attributes further includes:
- the entity and attribute joint model performs embedded training on the first type of triples and the second type of triples, and determines the first type of triples and the second type of triples
- the embedding vector of the triplet of the second type includes:
- the method before acquiring the input asset value query information for the user, the method further includes:
- An asset knowledge graph is constructed, wherein the asset knowledge graph includes multiple first-type triples in the format of entity-attribute-attribute value and multiple second-type triples in the format of head entity-relationship-tail entity.
- the method before acquiring the input asset value query information for the user, the method further includes:
- the asset knowledge graph determine a set of historical assets that have interaction events with the user, wherein the set of historical assets includes at least one historical asset;
- the graph convolutional network model is trained according to the asset value training sample, wherein the value of the asset in the asset value training sample to the user is determined according to the interaction event.
- the determining the value of the asset in the asset value training sample to the user according to the interaction event further includes:
- the interaction event and the path length determine the value of each of the historical assets in the historical asset set to the user, wherein the longer the path length, it indicates the value of the corresponding historical asset to the user The lower the value.
- the determining the value of each historical asset in the historical asset set to the user according to the interaction event and the path length includes:
- V A l-1 basevalue
- l represents the path length between the node where the user is located and the node where each historical asset is located in the historical asset collection
- basevalue represents the base value corresponding to the interaction event
- A is less than 1 constant.
- the input of the asset embedding vector of each of the assets into the graph convolutional network model to obtain the value of each of the assets to the user includes:
- the adjacent triplets include at least one triplet
- the spliced vector is input into a fully connected layer in the graph convolutional network model, and the value of the target asset to the user is output through the fully connected layer.
- the input of the concatenated vector into a fully connected layer in the graph convolutional network model, and outputting the value of the target asset to the user via the fully connected layer include:
- the value of the target asset to the user is obtained by using the following formula through the fully connected layer of the graph convolutional network model:
- ⁇ represents the sigmoid activation function
- C represents the final embedding vector of the user
- v represents the embedding vector of the target asset
- w represents the weight
- b represents the bias.
- the acquisition of the input asset value query information for the user includes:
- the input user information used to characterize the user and the set of assets to be evaluated are obtained, so that the graph convolutional network model outputs the value of each asset to be evaluated in the set of assets to be evaluated to the user.
- the acquisition of the input asset value query information for the user includes:
- the input user information used to characterize the user is acquired, so that the graph convolutional network model outputs at least one asset whose value to the user satisfies a preset value.
- the embodiment of the present disclosure also provides a model training method, which includes:
- the historical asset interaction information determine a historical asset set that has an interaction event with the user, wherein the historical asset set includes at least one historical asset;
- the training process based on the relationship between each of the historical assets and attributes includes:
- the training process based on the entities and attributes of each of the historical assets further includes:
- the method further includes:
- an asset value training sample for training the graph convolutional network model to be trained is constructed.
- the asset embedding vector of each of the historical assets is input into the graph convolutional network model to be trained, and the graph convolutional network model to be trained is trained to obtain the graph convolutional network model.
- Network models including:
- the adjacent triplets include at least one triplet
- n is positive integer
- the embedding vector of the target historical asset and the final embedding vector of the user are spliced through the splicing layer of the graph convolutional network model to be trained to obtain the spliced vector;
- the calculating the loss value according to the predicted asset value and the preset value of the target historical asset to the user includes:
- the loss value is calculated using the following formula:
- output represents the predicted asset value
- value represents the preset value
- ⁇ 1 and ⁇ 2 represent hyperparameters
- r represents the relationship embedding vector
- R represents the relationship matrix
- I represents the quality of triplets
- E represents the asset knowledge
- V represents the parameters of the graph convolutional network model to be trained.
- an asset value evaluation device which includes:
- the first memory is used to store computer programs
- the first processor is used to execute the computer program in the first memory to realize the following steps:
- the asset embedding vector of each of the assets is input into the graph convolutional network model to obtain the value of each of the assets to the user.
- the embodiment of the present disclosure also provides a model training device, which includes:
- the second memory is used to store computer programs
- the second processor is used to execute the computer program in the first memory to realize the following steps:
- the historical asset interaction information determine a historical asset set that has an interaction event with the user, wherein the historical asset set includes at least one historical asset;
- an embodiment of the present disclosure further provides a computer-readable storage medium, wherein:
- the readable storage medium stores computer instructions, and when the computer instructions are run on the computer, the computer is made to execute the asset value evaluation method described above or the model training method described above.
- FIG. 1 is a method flow chart of an asset value evaluation method provided by an embodiment of the present disclosure
- Fig. 2 is the method flowchart of the training process based on the relationship between the assets and attributes of each asset in step S103 in Fig. 1;
- Fig. 3 is the method flowchart of step S202 in Fig. 2;
- Fig. 4 is another method flowchart of the training process based on the relationship between the asset and the attribute of each asset in step S103 in Fig. 1;
- Fig. 5 is the method flowchart of step S402 in Fig. 4;
- FIG. 6 is a schematic structural diagram of an entity and attribute joint model in an asset value evaluation method provided by an embodiment of the present disclosure
- FIG. 7 is a flow chart of one of the methods before step S101 in FIG. 1;
- FIG. 8 is a flow chart of one of the methods after step S601 in FIG. 7;
- FIG. 9 is a schematic diagram of partial connections in an asset knowledge graph in an asset value evaluation method provided by an embodiment of the present disclosure.
- Fig. 10 is the method flowchart of step S103 in Fig. 1;
- FIG. 11 is a schematic structural diagram of a graph convolutional network model in an asset value evaluation method provided by an embodiment of the present disclosure
- FIG. 12 is a schematic structural diagram of one hop in an asset value evaluation method provided by an embodiment of the present disclosure.
- FIG. 13 is a method flowchart of a model training method provided by an embodiment of the present disclosure.
- Fig. 14 is a method flowchart of the training process based on the relationship between each historical asset and attribute in step S903 in Fig. 13;
- Fig. 15 is another method flowchart of the training process based on the relationship between each historical asset and attribute in step S903 in Fig. 13;
- Fig. 16 is a flow chart of the method after step S902;
- FIG. 17 is a flow chart of the method in step S904 in FIG. 13 .
- the embodiments of the present disclosure provide an asset value evaluation method, a model training method, a device, and a readable storage medium.
- FIG. 1 it is a method flowchart of an asset value evaluation method provided by an embodiment of the present disclosure.
- the asset value evaluation method can be applied to an asset value evaluation system, and the asset value evaluation system can realize the evaluation of intangible assets and tangible assets.
- Valuation of assets in which, intangible assets such as patents, papers, trademarks, etc., tangible assets such as mobile phones, laptops, air conditioners, TVs, etc.
- the method of asset valuation includes:
- the asset value query information when evaluating the value of a mobile phone, can be the model, serial number, etc. of the mobile phone; when evaluating the value of a patent, the asset value query information can also be the patent application number, IPC classification number, etc.; when the user is a company, the asset value query information can also include the user's identification number, such as the company organization code, company name, etc.; when the user is an individual, the asset value query information It may also include the user's name, ID number, etc.
- the asset value query information can also be set according to actual application needs, which is not limited here. The user may be an individual or a company, which is not limited here.
- step S101 obtain the input asset value query information for the user, including:
- the input user information used to characterize the user and the set of assets to be evaluated are obtained, so that the graph convolutional network model outputs the value of each asset to be evaluated in the set of assets to be evaluated to the user.
- the user information used to characterize the user and the set of assets to be evaluated can be input on the asset value evaluation system.
- the asset to be evaluated is a patent
- the patent can be input Patent application number, IPC classification number, etc.
- the asset to be evaluated is a mobile phone
- the user information that characterizes the user can be the organization Organization code, company name, etc.
- the user information characterizing the user may be a person's name, ID number, etc. In this way, not only the evaluation of personal asset value can be realized, but also the evaluation of company asset value can be realized.
- asset value of other objects can also be evaluated according to actual application needs, such as schools, families, etc.
- input user 1 and the set of assets to be evaluated including asset 1, asset 2 and asset 3 on the asset value evaluation system to determine the value of asset 1, asset 2 and asset 3 on user 1 respectively value.
- step S101 obtain the input asset value query information for the user, including:
- the input user information used to characterize the user is acquired, so that the graph convolutional network model outputs at least one asset whose value to the user satisfies a preset value.
- only the user information used to characterize the user may be input on the asset value evaluation system, for example, when the user is a company, the corresponding organization code may be input , company name, etc., when the user is an individual, it may be to input the corresponding name, ID number, etc.
- the organization code of user 2 is input into the asset value evaluation system to determine the value of each asset to user 2 from the historical asset collection in the asset value evaluation system.
- the number of assets to be screened can also be input on the asset value evaluation system.
- the assets of the corresponding number of assets that have been screened out for example, after inputting 100 assets to be When appraising assets for value evaluation, set the number of assets to be screened to 10. In this way, after evaluating the value of 100 assets to be evaluated, it is possible to select assets with higher value to users from the 100 assets to be evaluated. 10 assets, so as to achieve a specific number of asset recommendations for users.
- the user When it is determined that there is historical asset interaction information of the user, the user is a user who has historical interaction events such as application, purchase, and transfer of related assets, and it can be determined that the asset set is obtained by querying the asset value query information.
- the asset collection may include one asset, or multiple assets.
- the multiple assets may all be tangible assets, or all may be intangible assets, or may include both tangible assets and intangible assets, which is not limited here.
- S103 Make an embedding representation for each of the assets, and determine the asset embedding vector of each of the assets, wherein the asset embedding vector is obtained by training based on the relationship between each of the assets and attributes, and the attributes are used for Intrinsic parameters characterizing said asset;
- the asset embedding vector of each of the assets is obtained based on the relationship between each of the assets and attributes, and the attributes Intrinsic parameters used to characterize the asset. For example, if the asset is a patent, its corresponding attributes include the number of claims, number of times involved in litigation, patent publication number, patent application number, classification number, legal status, etc.; for another example, if the asset is a mobile phone, its corresponding attributes include Image resolution, screen size, etc.
- the asset embedding vector of each of the assets is trained based on the relationship between each of the assets and attributes, and the assets are entities in nature, the asset embedding vector not only incorporates the assets themselves The corresponding entity, and also integrates the relationship between the asset and the attributes of the asset itself. In this way, after the asset of the inquiring user is determined, not only the entity information of the asset can be determined, but also the attribute information related to the asset can be determined, so that the relevant information of the asset can be determined more accurately.
- S104 Input the asset embedding vector of each asset into the graph convolutional network model to obtain the value of each asset to the user.
- the graph convolutional network model may be a ripplenet (ripplenet) model, a knowledge graph convolutional network (KGCN) model, or a graph attention network (GAT) model, which is not limited here.
- the value of each asset to the user can be obtained by inputting the asset embedding vector of each asset into the graph convolutional network model. Since the asset embedding vector input into the graph convolutional network model is trained based on the relationship between the entities and attributes of each asset, the asset embedding vector incorporates the relationship between the asset and the attributes of the asset itself, After determining the asset of the querying user, not only the entity information of the asset can be determined, but also the attribute information related to the asset can be determined. After the asset embedding vector of each asset is input into the graph convolutional network model, it can be more accurate The connection between the asset and the user can be accurately determined, thereby improving the accuracy of asset value evaluation.
- step S103 the training process based on the relationship between each of the assets and attributes includes:
- S201 Determine the first type of triplet in the format of entity-attribute-attribute value in each asset
- S202 Initialize the embedding vector of the first type of triples and perform training to obtain the asset embedding vector.
- step S201 to step S202 is as follows:
- the format of each asset is a first-type triplet of entity-attribute-attribute value
- the format of the first-type triplet can be expressed as (h, a, a_value), where h represents an entity in the first type triple, a represents an attribute in the first type triple, and a_value represents an attribute value in the first type triple.
- the triplet (patent 1, legal status, public), “patent 1" means the entity, “legal status” means the attribute, and “public” means the attribute value; for another example, for the triplet (patent 1, claim number, 30), “patent 1” represents the entity, “claim number” represents the attribute, and “30” represents the “attribute value”; for another example, for the triplet (patent 1, number of times involved in litigation, 5), “patent 1 "Indicates an entity, “Number of Litigation Involved” indicates an attribute, and “5" indicates an “attribute value”.
- the first-type triplet of the asset may also be constructed according to actual application needs, and no further examples are given here.
- the first type of triples can be constructed based on the attributes that have a greater impact on the value of assets.
- the influence of value is relatively large.
- the attribute values corresponding to the attributes of the asset itself are fully considered, so that the value of the corresponding patent can be accurately determined.
- the construction of the first type of triples can improve the accuracy of asset value evaluation.
- the embedding vector of the first type of triples is initialized and trained to obtain the asset embedding vector. Since the constructed triplets of the first type fully consider the attribute values corresponding to the attributes of the asset itself, in this way, after initializing the embedding vectors of the triplets of the first type and performing training, the obtained The asset embedding vector fully considers the relationship between the entity corresponding to the asset and the inherent attributes of the entity itself. In this case, the graph convolutional network model can evaluate the asset value of the user from the two dimensions of entity and attribute, and improve The accuracy of asset valuation is improved.
- initializing the embedding vector of the triplet of the first type in step S202 includes:
- the first type of triples can be encoded through a bidirectional long-short-term memory (Bi-LSTM) network, and the entities in the first type of triples can be respectively encoded by the Bi-LSTM network , attributes, and attribute values are initialized and coded; it is also possible to use random initialization to embed entities and attributes in the first type of triples, and use a Bi-LSTM network to The attribute value of is initialized and coded, so as to obtain the initialization embedding vector of the first type of triplet.
- the first type of triples can be trained through the graph embedding conversion (Translate) model to obtain the corresponding asset embedding vectors.
- the graph embedding Translate model may be a TransH model, a TransR model, or a TransE model. It should be noted that embedding entities and attributes is essentially a process of converting text into vectors.
- S302 Encode the character set through the Bi-LSTM network, and take the result of the hidden layer in the Bi-LSTM network as the initialization code of the attribute value;
- step S301 to step S303 is as follows:
- the attribute value in the triplet of the first type split the attribute value in the triplet in the first type into multiple characters according to characters, and obtain the split character set, for example, the
- encode the character set through the Bi-LSTM network, and take the result of the hidden layer in the Bi-LSTM network as the initialization code of the attribute value, and then, according to the initialization code, determine the The initialization embedding vector of the attribute value.
- the entities and attributes of the first type of triples can be determined. Initialize the embedding vector, and then obtain the embedding vector of the first type of triplet according to the initialization embedding vector of entities, attributes and attribute values in the first type of triplet.
- the parameter values that is, attribute values
- the attribute value is the public number.
- the serial numbers of the same products sold in the same supermarket are often different.
- the attribute values in the first type of triples can be regarded as approximately infinite-dimensional, and infinite-dimensional is often difficult to train, and it is impossible to use the method of random initialization for training.
- the Bi-LSTM network encodes the character set split according to the characters in the first type of triplet.
- each character is a finite dimension, which can be Encoding avoids the problem of difficult initialization vectors.
- the relationship between the character and the preceding and following characters is fully considered, thus ensuring that the Bi-LSTM network
- the trained initialization vector can still express the meaning of the attribute value itself before character splitting, which ensures the accuracy of the asset embedding vector.
- step S103 the training process based on the relationship between each of the assets and attributes further includes:
- S401 Determine the second type of triplet in the format of head entity-relationship-tail entity in each asset
- S402 Perform embedded training on the triplets of the first type and the triplets of the second type through an entity and attribute joint model, and determine the triplets of the first type and the triplets of the second type Embedding vector.
- step S401 to step S402 is as follows:
- the format of the triplet of the second type can be expressed as (h, r, t), wherein, h represents the head entity in the triplet of the second type, and r represents the entity in the triplet of the second type relation, t represents the tail entity in the second type of triplet.
- h represents the head entity in the triplet of the second type
- r represents the entity in the triplet of the second type relation
- t represents the tail entity in the second type of triplet.
- the entities in the first type of triples It is the same entity as the entity in the second type of triplet.
- the triplets of the first type and the triplets of the second type are trained through the joint model of entities and attributes to obtain embedding vectors of corresponding triplets, and based on this, the asset embedding vectors of the assets can be determined.
- the embedding vector obtained after the entity and attribute joint model training fully integrates the relationship between assets and their own attributes; for the second type of triples, the The embedding vector obtained after the entity and attribute joint model training fully integrates the relationship between the asset and other entities, so that the The asset embedding vector not only integrates the relationship between assets and attributes, but also the relationship between assets and other entities, thus improving the accuracy of asset value evaluation based on asset embedding vectors.
- the first type of triples can be respectively determined
- the loss function of the group and the loss function of the second type of triples are further jointly trained according to the loss functions of the two parts.
- the specific joint training process can refer to the description of the relevant content below.
- the entity and attribute joint model may be a pre-built model based on the Bi_LSTM network and the graph embedding Translate model.
- step S402 perform embedded training on the first type of triples and the second type of triples through the entity and attribute joint model, and determine the first type of triples and the Embedding vectors for the second class of triples, consisting of:
- S501 Perform embedded training on the first type of triples based on the initialization embedding vector of the first type of triples through the first graph embedding conversion Translate model in the entity and attribute joint model;
- S502 Perform embedded training on the second type of triples through the second graph embedding conversion Translate model in the entity and attribute joint model;
- S503 Perform joint loss training based on the loss function of the first graph embedding transformation Translate model and the loss function of the second graph embedding transformation Translate model, to obtain the first type triplet and the second type triplet Embedding vectors for groups.
- step S501 the specific implementation process from step S501 to step S503 will be described in combination with the entity and attribute association model shown in Figure 6:
- the entities and attributes in the first type of triples may be embedded and represented by random initialization, and the embedding vectors of the entities and attributes in the first type of triples are respectively obtained, and then, through the The first graph embedding Translate model trains the embedding vectors of entities, attributes, and attribute values in the first type of triples.
- the loss function for training the first type of triples by embedding the Translate model in the first image is:
- the loss function for training the second type of triples is:
- ⁇ represents a positive sample triplet in the second type of triplet
- ⁇ ' represents a negative sample triplet
- the negative sample triplet is generated by randomly replacing the head entity or tail entity with the original triplet.
- the entities in the first type of triples and the second Entities in class triples are the same entity, such that the final loss function for training the joint model of entities and attributes can be expressed as:
- joint loss training may be performed based on the loss function of the first graph embedded in the Translate model and the loss function of the second graph embedded in the Translate model to obtain the first type of triplet and the second Embedding vectors for class triples.
- the first type of triples and the second type of triples are trained through the entity and attribute joint model, so that the obtained embedding vector of the first type of triples can better represent the
- the attribute value between the asset and its own attribute, the embedding vector of the second type of triple can better represent the relationship between the asset and other entities, for example, the head entity and tail entity where the asset is located
- the relationship between them ensures that the obtained asset embedding vector can better represent the relationship between the asset and its own attributes and the relationship between the asset and other entities, which provides a sufficient basis for the subsequent evaluation of asset value based on the asset embedding vector. protection.
- the first graph embedding Translate model and the second graph embedding Translate model may be a TransH model, a TransR model, or a TransE model.
- the acquisition of the first-type triples and the second-type triples can be obtained according to a pre-built asset knowledge map, specifically, in step S101: acquire the input user-specific assets Before value query information, the method also includes:
- An asset knowledge graph is constructed, wherein the asset knowledge graph includes multiple first-type triples in the format of entity-attribute-attribute value and multiple second-type triples in the format of head entity-relationship-tail entity.
- the correspondence table between the entities corresponding to the assets and the attributes of the entities that affect the value of the assets can be as shown in Table 1.
- asset attributes can also be set according to actual application needs, which is not limited here.
- a preset composition strategy is designed. Through the preset composition strategy, entities and the attributes corresponding to the entities can be added or deleted according to actual needs, thus ensuring the flexibility of building the asset knowledge map. . Then, analyze the data in the preset asset database according to the preset composition strategy, and obtain the asset data after analysis.
- the asset knowledge graph includes multiple first-type triples in the format of entity-attribute-attribute value and multiple second-type triples in the format of head entity-relationship-tail entity.
- the specific number of triples of the first type and the specific number of triples of the second type may be determined according to practical applications, and are not limited here.
- step S101 before estimating the value of the asset through the graph convolutional network model, it is also necessary to pre-build asset value training samples to train the graph convolutional network model to ensure asset value evaluation
- the accuracy of the result as shown in Figure 7, before step S101: obtaining the input query information for the user's asset value, the method further includes:
- S601 According to the asset knowledge graph, determine a historical asset set that has an interaction event with the user, wherein the historical asset set includes at least one historical asset;
- S602 Construct asset value training samples according to the historical asset collection
- step S601 to step S603 is as follows:
- the historical asset interaction information of the user may be by collecting the historical interaction records of the user on the asset value evaluation system, and according to the historical interaction records Determine the historical asset interaction information of the user, and then, according to the historical asset interaction information and the asset knowledge graph, determine a historical asset set that has an interaction event with the user, and the historical asset set includes at least one historical asset Assets, where the interaction event may be application, purchase, litigation, etc. Then, an asset value training sample is constructed according to the historical asset set, and the value of the asset to the user in the asset value training sample is determined according to the interaction event.
- the value corresponding to the interaction event "apply” is “8” points
- the value corresponding to the interaction event "purchase” is “10” points
- the value corresponding to the interaction event "involved and won the lawsuit” is “10” points .
- (P,C,V) ⁇ S ⁇ , i 1,2,...,n, where P represents historical assets, C represents users, and V represents value; the format of the asset value training sample can also be user-historical asset-value, corresponding to
- (C,P,V) ⁇ S ⁇ , i 1,2,...,n, where No limit. Then, the graph convolutional network model is trained according to the asset value training samples, so that the trained graph convolutional network model can be used to evaluate the value of assets to users accordingly.
- the value of the asset to the user may be determined according to the correspondence between the interaction event and the value of the asset.
- Value for example, for the case where the asset is a patent, three indicators can be used to evaluate the value of the patent: It is preset that the asset value when the interaction event is "apply” is “8" points, the asset value when the interaction event is “purchase” is “10” points, and the asset value when the interaction event is "involved and won the lawsuit" is " 10".
- step S601 after determining the historical asset set that has an interaction event with the user, the The method also includes:
- S701 Determine the path length between the node where the user is located and the node where each historical asset is located in the historical asset set in the asset knowledge graph;
- S702 According to the interaction event and the path length, determine the value of each of the historical assets in the historical asset set to the user, wherein the longer the path length, the corresponding historical asset is more valuable to the user. The lower the value of the user.
- step S701 to step S702 is as follows:
- the value of the asset to the user can be determined according to the interaction event and the path length.
- step S702 according to the interaction event and the corresponding relationship between the path length and asset value, determine the value of each of the historical assets in the historical asset set to the user, including :
- V A l-1 basevalue
- l represents the path length between the node where the user is located and the node where each historical asset is located in the historical asset set
- l is a positive integer
- basevalue represents the base value
- A is a constant less than 1 .
- the node where the user is located is directly connected to the node where any target historical asset in the historical asset set is located, correspondingly, the node where the user is located
- the path length between the node and the node where any target historical asset in the historical asset set is located is 1.
- the specific value of A may be 0.8 or 0.6, which is not limited here.
- step S103 Input the asset embedding vector of each asset into the graph convolutional network model to obtain the value of each asset to the user, including:
- S801 Determine the similarity between the embedding vector of any target asset in the asset set and the head entity in the adjacent triplet in the relational space, wherein the higher the similarity, the higher the similarity in the adjacent triplet The closer the head entity is to the target asset, the adjacent triplets include at least one triplet;
- S802 Propagate to the tail entities of the adjacent triples with the similarity as the weight, and after n times of propagation of the graph convolutional network model, obtain the final embedding vector of the user, where n is positive integer;
- S803 Splicing the embedding vector of the target asset and the final embedding vector of the user through the splicing layer of the graph convolutional network model to obtain a spliced vector;
- the graph convolutional network model may be a model based on the ripplenet network.
- the specific implementation process from step S801 to step S804 is explained:
- the adjacent triplets include at least one triplet. Taking the target asset as the head entity h, and starting from the head entity h to search for the relationship r and the tail entity t, such a complete triple is called a hop.
- the adjacent triplets of asset 1 (that is, the initial hop triplet) include four triplets, which are asset 1-relationship 1-entity 1, asset 1-relationship 2 -entity 2, asset 1-relationship 3-entity 3, asset 1-attribute 1-attribute value 1, when the next hop triplet starts, use the tail entity of the previous hop triplet as the next hop triplet
- the head entity which is continuously calculated, implements the "propagation" mentioned below.
- v represents the embedding vector of the target asset
- R and h represent the relation in the adjacent triplet and the embedding vector representation of the head entity, respectively.
- the user's final embedding vector is obtained, and n is a positive integer .
- n is a positive integer .
- n may be 2.
- the value of n may also be set according to actual needs, which is not limited here.
- the embedding vector of the target asset and the final embedding vector of the user are spliced through the splicing layer of the graph convolutional network model to obtain a spliced vector, which can be expressed as: concat( C, v); the spliced vector combines the embedding vector of the target asset and the final embedding vector of the user, ensuring that the information between the target asset and the user is fully fused, and the A graph convolutional network model for a targeted assessment of the value of the target asset to the user.
- the spliced vector is input into the fully connected layer in the graph convolutional network model, and the value of the target asset to the user is output through the fully connected layer, thereby passing through the graph convolutional network model An assessment of the value of the target asset is achieved.
- the mapping relationship between the concatenated vector and the value of the target asset to the user is determined, and based on the mapping relationship, the value of the target asset to the user can be accurately determined value, thereby improving the accuracy of asset value assessment for users.
- step S804 input the concatenated vector into the fully connected layer in the graph convolutional network model, and pass through the fully connected layer Outputting the value of the target asset to the user, including:
- the value of the target asset to the user is obtained by using the following formula through the fully connected layer of the graph convolutional network model:
- ⁇ represents the sigmoid activation function
- C represents the final embedding vector of the user
- v represents the embedding vector of the target asset
- w represents the weight
- b represents the bias.
- the embedding vector of the asset is determined based on the relationship between the asset and its own attributes, as well as the relationship between the asset and other entities, the accuracy of asset value evaluation for users is improved. Spend.
- an embodiment of the present disclosure also provides a model training method, which can be applied to an asset value evaluation system, wherein the model training method includes:
- S902 According to the historical asset interaction information, determine a historical asset set that has an interaction event with the user, where the historical asset set includes at least one historical asset;
- S903 Make an embedding representation for each of the historical assets, and determine the asset embedding vector of each of the historical assets, wherein the asset embedding vector of each of the historical assets is trained based on the relationship between each of the historical assets and attributes Obtained, the attributes are used to characterize the inherent parameters of the historical assets;
- S904 Input the asset embedding vector of each historical asset into the graph convolutional network model to be trained, and train the graph convolutional network model to obtain the graph convolutional network model.
- step S901 to step S904 is as follows:
- the user may be by collecting the historical interaction records of the user on the self-asset value evaluation system, determining the historical asset interaction information of the user according to the historical interaction records, and then determining the interaction with the user according to the historical asset interaction information.
- the user has a historical asset collection of interactive events, the historical asset collection includes at least one historical asset, where the interactive event can be application, purchase, litigation, transfer, etc., and then, each of the historical asset collection
- the historical assets are represented by embedding, so as to determine the asset embedding vector of each of the historical assets.
- the Attributes are used to characterize the inherent parameters of the historical asset, and the asset embedding vector of the historical asset not only integrates the entity corresponding to the asset itself, but also integrates the relationship between the entity and the attribute of the entity itself. Then, the asset embedding vector corresponding to each of the historical assets is input into the graph convolutional network model to be trained, and the graph convolutional network model to be trained is trained to obtain the graph convolutional network model.
- the asset embedding vector input into the graph convolutional network model to be trained is obtained based on the relationship between each of the historical assets and attributes
- the asset embedding vector of the historical asset combines the historical assets with the The relationship between the attributes of the historical assets themselves, when using the asset embedding vector of the historical assets to train the graph convolutional network model to be trained, the relevant information of the historical assets can be determined more accurately, so that Improved training accuracy for models used for asset valuation.
- step S903 the training process based on the relationship between each of the historical assets and attributes includes:
- S1001 Determine the first type of triplet in the format of entity-attribute-attribute value in each of the historical assets;
- S1002 Initialize the embedding vector of the first type of triples and perform training to obtain the asset embedding vector of the historical asset.
- step S1001 to step S1003 is as follows:
- the format of each of the historical assets is a first-type triplet of entity-attribute-attribute value
- the format of the first-type triplet can be expressed as (h, a, a_value), where, h represents an entity in the first-type triple, a represents an attribute in the first-type triple, and a_value represents an attribute value in the first-type triple.
- the first type of triples can be constructed based on the attributes that have a greater impact on the value of assets.
- the influence of value is relatively large.
- the attribute values corresponding to the attributes of the historical assets themselves are fully considered, so that the value of the corresponding patent can be accurately determined.
- the accuracy of training the model for asset value evaluation can be improved.
- the constructed triplet of the first type fully considers the attribute value corresponding to the attribute of the entity itself, in this way, after the embedding vector of the triplet of the first type is initialized and trained, the obtained The asset embedding vector of the historical asset fully considers the relationship between the historical asset and the inherent attributes of the historical asset itself.
- the graph convolutional network model to be trained Training improves the accuracy of training models for asset valuation.
- step S903 the training process based on the relationship between each of the historical assets and attributes further includes:
- S1101 Determine the second type of triplet in the format of head entity-relationship-tail entity in each of the historical assets;
- S1102 Perform embedded training on the triplets of the first type and the triplets of the second type through an entity and attribute joint model, and determine the triplets of the first type and the triplets of the second type Embedding vector. It should be noted that the embedded training of the first type of triples and the second type of triples is performed through the entity and attribute joint model, and the first type of triples and the second type of triples are determined. For the specific implementation process of the triplet-like embedding vector, you can refer to the description of the relevant part in the aforementioned asset value evaluation method, and will not repeat it here.
- step S902 after determining the set of historical assets that have interaction events with the user according to the historical asset interaction information, the method further includes:
- S1201 Determine the value of each historical asset to the user according to the interaction event
- S1202 According to the value of each historical asset to the user, construct an asset value training sample for training the graph convolutional network model to be trained.
- step S904 Input the asset embedding vector corresponding to each of the historical assets into the graph convolutional network model to be trained, and perform Training to obtain the graph convolutional network model, including:
- S1301 According to the pre-built asset knowledge graph, determine each of the historical assets that have interaction events with the user in the asset value training sample;
- S1302 Determine the similarity in the relational space between the embedding vector of any target historical asset in each of the historical assets and the head entity in the adjacent triplet, wherein the higher the similarity, the more the adjacent triplet The closer the head entity in the group is to the target historical asset, the adjacent triplets include at least one triplet;
- S1304 Splicing the embedding vector of the target historical asset and the final embedding vector of the user through the splicing layer of the graph convolutional network model to be trained to obtain a spliced vector;
- S1305 Input the concatenated vector into a fully connected layer in the graph convolutional network model to be trained, and output the predicted asset value of the target historical asset to the user through the fully connected layer;
- S1306 Calculate a loss value according to the predicted asset value and the preset value of the target historical asset to the user;
- S1307 Use the loss value to update the parameters of the graph convolutional network model to be trained to obtain the graph convolutional network model.
- v represents the embedding vector of the target historical asset
- R and h represent the relation in the adjacent triplet and the embedding vector representation of the head entity, respectively.
- the final embedding vector of the user After n times of propagation of the graph convolutional network model to be trained, the final embedding vector of the user obtained, n is a positive integer; the final embedding vector of the user is:
- ⁇ represents the sigmoid activation function
- C represents the final embedding vector of the user
- v represents the embedding vector of the target historical asset
- w represents the weight
- b represents the bias.
- step S1206 calculating the loss value according to the predicted asset value and the preset value of the target historical asset to the user, including:
- the loss value is calculated using the following formula:
- output represents the predicted asset value
- value represents the preset value
- ⁇ 1 and ⁇ 2 represent hyperparameters
- r represents the relationship embedding vector
- R represents the relationship matrix
- I represents the triplet is good or bad
- E represents the asset knowledge
- V represents the parameters of the graph convolutional network model to be trained.
- the aforementioned graph convolutional network model to be trained is trained and the trained graph convolutional network model is obtained, after inputting the asset value query information for the user, the corresponding asset set can be input into In the graph convolutional network model, the value of each asset in the asset collection is returned, thereby realizing the personalized value evaluation for different users.
- the user's personalized asset value evaluation can be realized through online learning, that is, it is necessary to train an asset value evaluation model based on the corresponding user , thereby ensuring the accuracy of the asset value evaluation of the corresponding user.
- the online learning method is called to retrain the parameters of the graph convolutional network model to be trained, and then the asset value evaluation of the user’s related assets can be performed, thereby improving the asset value evaluation. the accuracy.
- the problem-solving principle of the model training method provided by the embodiments of the present disclosure is similar to that of the asset value evaluation method described above.
- the relevant implementation process can refer to the description of relevant parts in the asset value evaluation method described above, and the repetition will not be repeated.
- an embodiment of the present disclosure also provides an asset value evaluation device, which includes:
- the first memory is used to store computer programs
- the first processor is used to execute the computer program in the first memory to realize the following steps:
- the asset embedding vector of each of the assets is input into the graph convolutional network model to obtain the value of each of the assets to the user.
- an embodiment of the present disclosure also provides a model training device, which includes:
- the second memory is used to store computer programs
- the second processor is used to execute the computer program in the first memory to realize the following steps:
- the historical asset interaction information determine a historical asset set that has an interaction event with the user, wherein the historical asset set includes at least one historical asset;
- an embodiment of the present disclosure also provides a computer-readable storage medium, wherein:
- the readable storage medium stores computer instructions, and when the computer instructions are run on the computer, the computer is made to execute the asset value evaluation method described above or the model training method described above.
- the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Accordingly, the present disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
- computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
- These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions
- the device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computational Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Analysis (AREA)
- Databases & Information Systems (AREA)
- Algebra (AREA)
- Artificial Intelligence (AREA)
- Operations Research (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
本公开提供了一种资产价值评估方法、模型训练方法、装置及可读存储介质,其中,所述资产价值评估方法,包括:获取输入的针对用户的资产价值查询信息;在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
Description
本公开涉及计算机技术领域,特别涉及一种资产价值评估方法、模型训练方法、装置及可读存储介质。
随着市场经济的发展,诸如商品、专利等资产的交易活动愈加频繁。如何实现对资产价值的评估就变得格外重要。
发明内容
本公开提供了一种资产价值评估方法、模型训练方法、装置及可读存储介质,用于提高资产价值评估的精确度。
第一方面,本公开实施例提供了一种资产价值评估方法,其中,包括:
获取输入的针对用户的资产价值查询信息;
在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;
对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;
将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
在一种可能的实现方式中,所述基于每个所述资产和属性的关系的训练过程,包括:
初始化所述第一类三元组的嵌入向量并进行训练,得到所述资产嵌入向量。
在一种可能的实现方式中,所述初始化所述第一类三元组的嵌入向量, 包括:
通过Bi-LSTM网络对所述第一类三元组进行编码,得到所述第一类三元组的初始化嵌入向量。
在一种可能的实现方式中,所述基于每个所述资产和属性的关系的训练过程,还包括:
确定每个所述资产中的格式为头实体-关系-尾实体的第二类三元组;
通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量每。
在一种可能的实现方式中,所述通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量,包括:
通过所述实体和属性联合模型中的第一图嵌入Translate模型基于所述第一类三元组的初始化嵌入向量对所述第一类三元组做嵌入式训练;
基于所述第一图嵌入Translate模型的损失函数与所述第二图嵌入Translate模型的损失函数进行联合损失训练,得到所述第一类三元组和所述第二类三元组的嵌入向量。
在一种可能的实现方式中,在所述获取输入的针对用户的资产价值查询信息之前,所述方法还包括:
构建资产知识图谱,其中,所述资产知识图谱包括多个格式为实体-属性-属性值的第一类三元组和多个格式为头实体-关系-尾实体的第二类三元组。
在一种可能的实现方式中,在所述获取输入的针对用户的资产价值查询信息之前,所述方法还包括:
根据所述资产知识图谱,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;
根据所述历史资产集合,构建资产价值训练样本;
根据所述资产价值训练样本对所述图卷积网络模型进行训练,其中,所述资产价值训练样本中的资产对所述用户的价值是根据所述交互事件所确定 的。
在一种可能的实现方式中,所述根据所述交互事件确定所述资产价值训练样本中的资产对所述用户的价值,还包括:
确定所述资产知识图谱中,所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度;
根据所述交互事件以及所述路径长度,确定所述历史资产集合中每个所述历史资产对所述用户的价值,其中,所述路径长度越长,表明相应的历史资产对所述用户的价值越低。
在一种可能的实现方式中,所述根据所述交互事件以及所述路径长度,确定所述历史资产集合中每个所述历史资产对所述用户的价值,包括:
采用以下公式确定所述历史资产集合中每个所述历史资产对所述用户的价值:
V=A
l-1basevalue
其中,l表示所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度,basevalue表示所述交互事件对应的基础价值,A为小于1的常数。
在一种可能的实现方式中,所述将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值,包括:
确定所述资产集合中任一目标资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标资产的联系越紧密,所述相邻三元组包括至少一个三元组;
以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;
通过所述图卷积网络模型的拼接层将所述目标资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;
将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值。
在一种可能的实现方式中,所述将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值,包括:
通过所述图卷积网络模型的全连接层采用以下公式获得所述目标资产对所述用户的价值:
y=σ(concat(C,v)·ω+b)
其中,σ表示sigmoid激活函数,C表示所述用户的最终嵌入向量,v表示所述目标资产的嵌入向量,w表示权重,b表示偏置。
在一种可能的实现方式中,所述获取输入的针对用户的资产价值查询信息,包括:
获取输入的用于表征所述用户的用户信息,以及待评估资产集合,以使所述图卷积网络模型输出所述待评估资产集合中每个待评估资产对所述用户的价值。
在一种可能的实现方式中,所述获取输入的针对用户的资产价值查询信息,包括:
获取输入的用于表征所述用户的用户信息,以使所述图卷积网络模型输出对所述用户的价值满足预设价值的至少一个资产。
第二方面,本公开实施例还提供了一种模型训练方法,其中,包括:
获取用户的历史资产交互信息;
根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;
对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到的,所述属性用于表征所述历史资产的固有参数;
将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
在一种可能的实现方式中,所述基于每个所述历史资产和属性的关系的 训练过程,包括:
确定每个所述历史资产中的格式为实体-属性-属性值的第一类三元组;
初始化所述第一类三元组的嵌入向量并进行训练,得到所述历史资产的资产嵌入向量。
在一种可能的实现方式中,所述基于每个所述历史资产的实体和属性的训练过程,还包括:
确定每个所述历史资产中的格式为头实体-关系-尾实体的第二类三元组;
通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。
在一种可能的实现方式中,在所述根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合之后,所述方法还包括:
根据所述交互事件,确定每个所述历史资产对所述用户的价值;
根据每个所述历史资产对所述用户的价值,构建对所述待训练的图卷积网络模型进行训练的资产价值训练样本。
在一种可能的实现方式中,所述将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型,包括:
根据预先构建的资产知识图谱,确定所述资产价值训练样本中与所述用户存在交互事件的各个所述历史资产;
确定各个所述历史资产中任一目标历史资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标历史资产的联系越紧密,所述相邻三元组包括至少一个三元组;
以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述待训练的图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;
通过所述待训练的图卷积网络模型的拼接层将所述目标历史资产的嵌入 向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;
将所述拼接后的向量输入所述待训练的图卷积网络模型中的全连接层,经所述全连接层输出所述目标历史资产对所述用户的预测资产价值;
根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值;
利用所述损失值对所述待训练的图卷积网络模型进行参数更新,获得所述图卷积网络模型。
在一种可能的实现方式中,所述根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值,包括:
采用以下公式计算损失值:
其中,output表示所述预测资产价值,value表示所述预设价值,λ1和λ2表示超参数,r表示关系嵌入向量,R表示关系矩阵,I表示三元组好坏,E表示所述资产知识图谱中的实体矩阵,V表示所述待训练的图卷积网络模型的参数。
第三方面,本公开实施例还提供了一种资产价值评估装置,其中,包括:
第一存储器和第一处理器;
其中,所述第一存储器用于存储计算机程序;
所述第一处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:
获取输入的针对用户的资产价值查询信息;
在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;
对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;
将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
第四方面,本公开实施例还提供了一种模型训练装置,其中,包括:
第二存储器和第二处理器;
其中,所述第二存储器用于存储计算机程序;
所述第二处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:
获取用户的历史资产交互信息;
根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;
对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到,所述属性用于表征所述历史资产的固有参数;
将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
第五方面,本公开实施例还提供了一种计算机可读存储介质,其中:
所述可读存储介质存储有计算机指令,当计算机指令在计算机上运行时,使得计算机执行如上面任一项所述的资产价值评估方法或者如上面任一项所述的模型训练方法。
图1为本公开实施例提供的一种资产价值评估方法的方法流程图;
图2为图1中步骤S103中基于每个资产的资产和属性的关系的训练过程的方法流程图;
图3为图2中步骤S202的方法流程图;
图4为图1中步骤S103中基于每个资产的资产和属性的关系的训练过程的另外一种方法流程图;
图5为图4中步骤S402的方法流程图;
图6为本公开实施例提供的一种资产价值评估方法中的实体和属性联合模型的其中一种结构示意图;
图7为在图1中步骤S101之前的其中一种方法流程图;
图8为在图7中步骤S601之后的其中一种方法流程图;
图9为本公开实施例提供的一种资产价值评估方法中资产知识图谱中的部分连接示意图;
图10为图1中步骤S103的方法流程图;
图11为本公开实施例提供的一种资产价值评估方法中的图卷积网络模型的其中一种结构示意图;
图12为本公开实施例提供的一种资产价值评估方法中的其中一个hop的结构示意图;
图13为本公开实施例提供的一种模型训练方法的方法流程图;
图14为图13中步骤S903中基于每个历史资产和属性的关系的训练过程的方法流程图;
图15为在图13中步骤S903中基于每个历史资产和属性的关系的训练过程的另外一种方法流程图;
图16为在步骤S902之后的方法流程图;
图17为图13中步骤S904的方法流程图。
为使本公开实施例的目的、技术方案和优点更加清楚,下面将结合本公开实施例的附图,对本公开实施例的技术方案进行清楚、完整地描述。显然,所描述的实施例是本公开的一部分实施例,而不是全部的实施例。并且在不冲突的情况下,本公开中的实施例及实施例中的特征可以相互组合。基于所描述的本公开的实施例,本领域普通技术人员在无需创造性劳动的前提下所获得的所有其他实施例,都属于本公开保护的范围。
除非另外定义,本公开使用的技术术语或者科学术语应当为本公开所属领域内具有一般技能的人士所理解的通常意义。本公开中使用的“包括”或者“包含”等类似的词语意指出现该词前面的元件或者物件涵盖出现在该词后面列举的元件或者物件及其等同,而不排除其他元件或者物件。
需要注意的是,附图中各图形的尺寸和形状不反映真实比例,目的只是示意说明本公开内容。并且自始至终相同或类似的标号表示相同或类似的元件或具有相同或类似功能的元件。
在相关技术中,主要基于固定指标和维度,结合人工或机器学习为诸如商品、专利等资产进行价值评估,然而,人工方式带有主观性,而且固定指标和维度无法为不同的用户做个性化定值,甚至无法体现指标之间的关系,价值评估结果较差。可见,现有存在资产价值评估精确度较低的技术问题。
鉴于此,本公开实施例提供了一种资产价值评估方法、模型训练方法、装置及可读存储介质。
如图1所示为本公开实施例提供的一种资产价值评估方法的方法流程图,该资产价值评估方法可以应用于资产价值评估系统中,通过该资产价值评估系统可以实现对无形资产和有形资产的价值评估,其中,无形资产诸如专利、论文、商标等,有形资产诸如手机、笔记本电脑、空调、电视等,该资产价值评估方法包括:
S101:获取输入的针对用户的资产价值查询信息;
比如,在对手机进行价值评估时,所述资产价值查询信息可以是该手机的型号、序列号等;在对专利进行价值评估时,所述资产价值查询信息还可以是该专利的申请号、IPC分类号等;在用户为公司时,所述资产价值查询信息还可以包括该用户的身份标识号,比如,公司组织机构代码、公司名称等;在用户为个人时,所述资产价值查询信息还可以包括该用户的人名、身份证号等。当然,还可以根据实际应用需要来设置所述资产价值查询信息,在此不做限定。所述用户可以是个人,还可以是公司,在此不做限定。
在具体实施过程中,步骤S101:获取输入的针对用户的资产价值查询信 息,包括:
获取输入的用于表征所述用户的用户信息,以及待评估资产集合,以使所述图卷积网络模型输出所述待评估资产集合中每个待评估资产对所述用户的价值。
在本公开实施例的其中一种应用场景中,可以在资产价值评估系统上输入用于表征所述用户的用户信息以及待评估资产集合,在待评估资产为专利时,可以是输入该专利的专利申请号、IPC分类号等,在待评估资产为手机时,还可以是输入该手机的型号、序列号等,其中,在所述用户为公司时,表征所述用户的用户信息可以是组织机构代码、公司名称等,在所述用户为个人时,表征所述用户的用户信息可以是人名、身份证号等。如此一来,不仅可以实现对个人资产价值的评估,还可以实现对公司资产价值的评估,当然,还可以根据实际应用需要来对其他对象资产价值的评估,比如,学校、家庭等,在此不做限定。举个具体的例子来说,在资产价值评估系统上输入用户1以及包括资产1、资产2和资产3在内的待评估资产集合,以确定资产1、资产2和资产3分别对用户1的价值。
在具体实施过程中,步骤S101:获取输入的针对用户的资产价值查询信息,包括:
获取输入的用于表征所述用户的用户信息,以使所述图卷积网络模型输出对所述用户的价值满足预设价值的至少一个资产。
在本公开实施例的另外一种应用场景中,可以仅在资产价值评估系统上输入用于表征所述用户的用户信息,比如,在所述用户为公司时,可以是输入相应的组织机构代码、公司名称等,在所述用户为个人时,可以是输入相应的人名、身份证号等。举个具体的例子来说,在资产价值评估系统上输入用户2的组织机构代码,以从资产价值评估系统的历史资产集合中,确定出各个资产对用户2的价值。此外,还可以在资产价值评估系统上输入所需筛选出的资产数量,在确定出各个资产对所述用户的价值之后,可以将筛选出的相应资产数量的资产,比如,在输入100个待评估资产进行价值评估时, 设定所需筛选出的资产数量为10,如此一来,在对100个待评估资产进行价值评估后,可以从100个待评估资产中筛选出对用户价值较高的10个资产,从而实现了对用户进行特定数量的资产推荐。
当然,除了上述提及的应用场景之外,还可以将本公开实施例中资产价值评估方法应用于其它的场景中,在此不再详述。
S102:在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;
在确定存在所述用户的历史资产交互信息时,所述用户为存在对相关资产诸如申请、购买、转让等历史交互事件的用户,可以确定采用所述资产价值查询信息查询得到资产集合,所述资产集合可以包括一个资产,还可以包括多个资产,多个资产可以均是有形资产,还可以均是无形资产,还可以是既包括有形资产又包括无形资产,在此不做限定。
S103:对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;
对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,每个所述资产的所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数。比如,所述资产为专利,其相应的属性包括权利要求数,涉诉次数、专利公开号、专利申请号、分类号、法律状态等;再比如,所述资产为手机,其相应的属性包括图像分辨率、屏幕尺寸等。由于每个所述资产的所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,而所述资产本质上为实体,如此一来,所述资产嵌入向量不仅融合了资产自身对应的实体,而且还融合了资产与资产自身的属性之间的关系。这样的话,当确定查询的用户的资产之后,不仅可以确定该资产的实体信息,还可以确定与该资产相关的属性信息,从而可以更精确地确定资产的相关信息。
S104:将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每 个所述资产对所述用户的价值。
所述图卷积网络模型可以是水波网络(ripplenet)模型,还可以是知识图卷积网络(KGCN)模型,还可以是图注意力网络(GAT)模型,在此不做限定。将每个所述资产的资产嵌入向量输入图卷积网络模型,可以得到每个所述资产对所述用户的价值。由于输入至所述图卷积网络模型中的资产嵌入向量是基于每个所述资产的实体和属性的关系训练得到的,所述资产嵌入向量融合了资产与资产自身的属性之间的关系,当确定查询的用户的资产之后,不仅可以确定该资产的实体信息,还可以确定与该资产相关的属性信息,将每个所述资产的资产嵌入向量输入图卷积网络模型之后,可以更精确地确定资产和所述用户之间的联系,进而提高了资产价值评估的精确度。
在本公开实施例中,为了实现对每个所述资产的资产嵌入向量的确定,如图2所示,步骤S103中:基于每个所述资产和属性的关系的训练过程,包括:
S201:确定每个所述资产中的格式为实体-属性-属性值的第一类三元组;
S202:初始化所述第一类三元组的嵌入向量并进行训练,得到所述资产嵌入向量。
在具体实施过程中,步骤S201至步骤S202的具体实现过程如下:
首先,确定每个所述资产中的格式为实体-属性-属性值的第一类三元组,所述第一类三元组的格式可以表示为(h,a,a_value),其中,h表示所述第一类三元组中的实体,a表示所述第一类三元组中的属性,a_value表示所述第一类三元组中的属性值。比如,对于三元组(专利1,法律状态,公开),“专利1”表示实体,“法律状态”表示属性,“公开”表示属性值;再比如,对于三元组(专利1,权利要求数,30),“专利1”表示实体,“权利要求数”表示属性,“30”表示“属性值”;再比如,对于三元组(专利1,涉诉次数,5),“专利1”表示实体,“涉诉次数”表示属性,“5”表示“属性值”。当然,还可以根据实际应用需要构造所述资产的所述第一类三元组,在此不再一一举例说明。在具体实施过程中,可以基于对资产价值影响较大的属性来构造所 述第一类三元组,比如,专利的“法律状态”、“权利要求数”、“涉诉次数”往往对专利价值的影响程度较大,通过构造包括这些属性的第一类三元组,充分考虑了资产自身的属性所对应的属性值,从而能够精确地确定相应专利的价值,基于相同的实现原理,通过所述第一类三元组的构造,可以提高资产价值评估的精确度。
然后,初始化所述第一类三元组的嵌入向量并进行训练,得到所述资产嵌入向量。由于所构造的所述第一类三元组充分考虑了资产自身的属性所对应的属性值,如此一来,初始化所述第一类三元组的嵌入向量并进行训练之后,所得到的所述资产嵌入向量充分考虑了资产对应实体与实体自身固有属性之间的关系,这样的话,所述图卷积网络模型可以从实体和属性两个维度来进行所述用户的资产价值的评估,提高了资产价值评估的精确度。
在本公开实施例中,为了得到所述第一类三元组的初始化嵌入向量,步骤S202中初始化所述第一类三元组的嵌入向量,包括:
通过Bi-LSTM网络对所述第一类三元组进行编码,得到所述第一类三元组的初始化嵌入向量。
在具体实施过程中,可以通过双向长短时记忆(Bi-LSTM)网络对所述第一类三元组进行编码,可以是通过Bi-LSTM网络分别对所述第一类三元组中的实体、属性和属性值进行初始化编码;还可以是采用随机初始化的方式对所述第一类三元组中的实体和属性进行嵌入表示,通过Bi-LSTM网络对所述第一类三元组中的属性值进行初始化编码,从而得到所述第一类三元组的初始化嵌入向量。在得到所述第一类三元组的初始化嵌入向量之后,可以通过图嵌入转换(Translate)模型对所述第一类三元组进行训练,得到相应的资产嵌入向量,具体的训练过程可以参照下文部分中的说明。在本公开的一些实施例中,图嵌入Translate模型可以是TransH模型、TransR模型、TransE模型。需要说明的是,将实体和属性进行嵌入表示本质上为将文字转换为向量的过程。
在本公开实施例中,如图3所示,步骤:通过Bi-LSTM网络对所述第一 类三元组中的属性值进行初始化编码,得到所述第一类三元组的初始化嵌入向量,包括:
S301:将所述第一类三元组中的属性值按字符拆开,获得拆开后的字符集合;
S302:通过Bi-LSTM网络对所述字符集合编码,并取所述Bi-LSTM网络中的隐藏层的结果作为所述属性值的初始化编码;
S303:根据所述初始化编码,得到所述第一类三元组的初始化嵌入向量。
在具体实施过程中,步骤S301至步骤S303的具体实现过程如下:
首先,确定所述第一类三元组中的属性值,将所述第一类三元组中的属性值按照字符拆分为多个字符,获得拆分后的字符集合,比如,所述属性值包括m个字符,m为正整数,所述字符集合可以表示为:C={c
1,c
2,c
3,...,c
m}。然后,通过所述Bi-LSTM网络对所述字符集合编码,并取所述Bi-LSTM网络中的隐藏层的结果作为所述属性值的初始化编码,然后,根据所述初始化编码,确定所述属性值的初始化嵌入向量,此时,在采用随机初始化的方式对所述第一类三元组中的实体和属性进行嵌入表示之后,可以确定所述第一类三元组的实体和属性的初始化嵌入向量,然后,根据所述第一类三元组中的实体、属性和属性值的初始化嵌入向量,得到所述第一类三元组的嵌入向量。
需要说明的是,不同资产的自身属性对应的参数值(即属性值)可以是多种情况,比如,随着时间的推移,在后公开的专利将被分配不同于在先公开的专利的公开号,这样的话,对于属性值为公开号的情况,是无法穷举的,再比如,同一超市所售卖的同种商品,各个商品的序列号往往不同。也就是说,对于所述第一类三元组中的属性值可以视为近似于无限维,而无限维往往难以训练,无法使用随机初始化的方法来进行训练,在本公开实施例中,通过所述Bi-LSTM网络对所述第一类三元组中按照字符拆分后的字符集合进行编码,一方面,将属性值按照字符拆分后,每个字符为有限维,可对其进行编码,避免了难以初始化向量的问题,另一方面,通过Bi-LSTM网络在对单个字符进行嵌入表示过程中充分考虑了该字符与前后字符之间的关联关系, 从而保证了经过Bi-LSTM网络训练后的初始化向量仍能表达字符拆分前属性值本身的含义,保证了资产嵌入向量的精确度。
在本公开实施例中,为了进一步地提高资产嵌入向量的精确度,从而更精确地确定资产和所述用户之间的联系,进而提高了资产价值评估的精确度,如图4所示,对于步骤S103中:基于每个所述资产和属性的关系的训练过程,还包括:
S401:确定每个所述资产中的格式为头实体-关系-尾实体的第二类三元组;
S402:通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。
在具体实施过程中,步骤S401至步骤S402的具体实现过程如下:
在基于每个所述资产和属性的关系的训练过程中,除了需要确定每个所述资产的第一类三元组之外,还需要确定每个所述资产中的格式为头实体-关系-尾实体的第二类三元组;如此一来,在考虑所述资产与自身属性之间的关系的同时,将所述资产与其它实体之间的关系也考虑了进来,这样的话,在获得资产嵌入向量之后,可以得到更精确的价值评估。所述第二类三元组的格式可以表示为(h,r,t),其中,h表示所述第二类三元组中的头实体,r表示所述第二类三元组中的关系,t表示所述第二类三元组中的尾实体。比如,对于专利2,(专利2,属于,D公司)中,“专利2”和“D公司”表示实体,“属于”表示关系。
然后,通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。需要说明的是,在通过所述实体和属性联合模型对所述第一类三元组和所述第二类三元组进行联合训练的过程中,所述第一类三元组中的实体和所述第二类三元组中的实体为同一实体。所述第一类三元组和所述第二类三元组通过所述实体和属性联合模型训练,获得相应三元组的嵌入向量,基于此,可以确定所述资产的资产嵌入向量。对于所述第一类三元组,经所述实体和属性联合模型训练后得到的嵌入向量充分融合了资产与资产自身属性之间的关 系,对于所述第二类三元组,经所述实体和属性联合模型训练后得到的嵌入向量充分融合了所述资产与其它实体之间的关系,如此一来,经所述第一类三元组和所述第二类三元组所获得的资产嵌入向量既融合了资产与属性之间的关系,又融合了资产与其它实体之间的关系,从而提高了基于资产嵌入向量进行资产价值评估的精确度。
需要说明的是,在通过所述实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练的过程中,可以分别确定所述第一类三元组的损失函数,以及所述第二类三元组的损失函数,进一步地根据这两部分的损失函数进行联合训练,具体的联合训练过程可以参照下文中相关内容的描述。
在本公开实施例中,所述实体和属性联合模型可以是基于所述Bi_LSTM网络和图嵌入Translate模型所预先构建的模型。如图5所示,步骤S402:通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量,包括:
S501:通过所述实体和属性联合模型中的第一图嵌入转换Translate模型基于所述第一类三元组的初始化嵌入向量对所述第一类三元组做嵌入式训练;
S502:通过所述实体和属性联合模型中的第二图嵌入转换Translate模型对所述第二类三元组做嵌入式训练;
S503:基于所述第一图嵌入转换Translate模型的损失函数与所述第二图嵌入转换Translate模型的损失函数进行联合损失训练,得到所述第一类三元组和所述第二类三元组的嵌入向量。
在具体实施过程中,结合图6所示的实体和属性联合模型,对步骤S501至步骤S503的具体实现过程进行说明:
确定每个所述资产的资产嵌入向量之前,首先,通过所述实体和属性联合模型中的第一图嵌入Translate模型基于所述第一类三元组的初始化嵌入向量对所述第一类三元组做嵌入式训练;对于所述第一类三元组,可以在通过所述Bi-LSTM网络确定所述第一类三元组中的属性值的初始化编码之后,再 使用所述第一图嵌入Translate模型训练所述第一类三元组。其中,可以是采用随机初始化的方式对所述第一类三元组中的实体和属性进行嵌入表示,分别获得所述第一类三元组中的实体和属性的嵌入向量,然后,通过所述第一图嵌入Translate模型训练所述第一类三元组中的实体、属性和属性值的嵌入向量。其中,通过所述第一图嵌入Translate模型对所述第一类三元组进行训练的损失函数为:
还可以通过所述实体和属性联合模型中的第二图嵌入Translate模型对所述第二类三元组做嵌入式训练;对于所述第二类三元组,可以是直接通过所述第二图嵌入Translate模型来训练所述第一类三元组,具体的训练方法包括:
首先,分别将所述第二类三元组中的头实体和尾实体映射到一个超平面w上,获得映射后的头实体h
⊥和映射后的尾实体t
⊥,如下式表示:
然后,使映射后的头实体向量加关系向量更接近于尾实体向量,获得所述第二类三元组的差值表示:
对所述第二类三元组进行训练的损失函数为:
其中,Δ表示所述第二类三元组中的正样本三元组,Δ'表示负样本三元组,所述负样本三元组由原三元组随机替换头实体或尾实体生成。
此外,所述第一类三元组中的实体与所述第二类三元组中的实体的相似度为:
在通过所述实体和属性联合模型对所述第一类三元组和所述第二类三元组进行联合训练的过程中,所述第一类三元组中的实体与所述第二类三元组 中的实体是相同的实体,如此一来,对所述实体和属性联合模型的训练的最终损失函数可以表示为:
Loss=Loss(e)+Loss(a)+Loss(sim)
在具体实施过程中,可以基于所述第一图嵌入Translate模型的损失函数与所述第二图嵌入Translate模型的损失函数进行联合损失训练,得到所述第一类三元组和所述第二类三元组的嵌入向量。通过所述实体和属性联合模型对所述第一类三元组和所述第二类三元组进行训练,使得所得的所述第一类三元组的嵌入向量能够更好地表示所述资产与自身属性之间的属性值,所述第二类三元组的嵌入向量能够更好地表示所述资产与其它实体之间的关系,比如,所述资产所处的头实体与尾实体之间的关系,从而保证了所得的资产嵌入向量能够更好地表示资产与资产自身属性之间的关系以及资产与其它实体之间的关系,为后续基于资产嵌入向量进行资产价值评估提供了充分的保障。
在本公开的一些实施例中,所述第一图嵌入Translate模型和所述第二图嵌入Translate模型可以是TransH模型、TransR模型、TransE模型。
在本公开实施例中,所述第一类三元组和所述第二类三元组的获取可以根据预先构建的资产知识图谱得到,具体地,在步骤S101:获取输入的针对用户的资产价值查询信息之前,所述方法还包括:
构建资产知识图谱,其中,所述资产知识图谱包括多个格式为实体-属性-属性值的第一类三元组和多个格式为头实体-关系-尾实体的第二类三元组。
对于所述资产知识图谱的构建过程,可以是根据先验经验,确定影响资产价值的资产属性。以资产为专利为例,资产对应的实体以及影响资产价值的实体的属性之间的对应关系表可以是如表1所示。当然,还可以根据实际应用需要来设置资产属性,在此不做限定。然后,根据所述资产属性,设计预设构图策略,通过所述预设构图策略,可以根据实际需要添加或者删除实体以及该实体所对应的属性,从而保证了所述资产知识图谱构建的灵活性。然后,根据所述预设构图策略对预设资产数据库中的数据进行解析,获得解 析后的资产数据。然后,可以将所述解析后的资产数据导入图数据库中,所述图数据库可以是neo4j,从而实现了对所述资产知识图谱的构建。所述资产知识图谱包括多个格式为实体-属性-属性值的第一类三元组和多个格式为头实体-关系-尾实体的第二类三元组。所述第一类三元组的具体个数,以及所述第二类三元组的具体个数可以根据实应用来确定,在此不做限定。
表1
在本公开实施例中,在通过所述图卷积网络模型预估所述资产的价值之前,还需要预先构建资产价值训练样本来对所述图卷积网络模型进行训练,以保证资产价值评估结果的精确度,如图7所示,在步骤S101:获取输入的针对用户的资产价值查询信息之前,所述方法还包括:
S601:根据所述资产知识图谱,确定与所述用户存在交互事件的历史资 产集合,其中,所述历史资产集合包括至少一个历史资产;
S602:根据所述历史资产集合,构建资产价值训练样本;
S603:根据所述资产价值训练样本对所述图卷积网络模型进行训练,其中,所述资产价值训练样本中的资产对所述用户的价值是根据所述交互事件所确定的。
在具体实施过程中,步骤S601至步骤S603的具体实现过程如下:
在获取输入的针对用户的资产价值查询信息之前,可以是获取所述用户的所述历史资产交互信息,可以是通过收集所述用户在资产价值评估系统上的历史交互记录,根据该历史交互记录确定所述用户的所述历史资产交互信息,然后,根据所述历史资产交互信息和所述资产知识图谱,确定与所述用户存在交互事件的历史资产集合,所述历史资产集合包括至少一个历史资产,其中,所述交互事件可以是申请、购买、涉诉等。然后,根据所述历史资产集合,构建资产价值训练样本,所述资产价值训练样本中的资产对所述用户的价值是根据所述交互事件所确定的。比如,交互事件“申请”对应的价值为“8”分,交互事件为“购买”时对应的价值为“10”分,交互事件为“涉诉且胜诉”时对应的价值为“10”分。所述资产价值训练样本的格式可以是历史资产-用户-价值,对应的训练样本集合可以表示为s
i={(P
i,C
i,V
i)|(P,C,V)∈S},i=1,2,......,n,其中,P表示历史资产,C表示用户,V表示价值;所述资产价值训练样本的格式还可以是用户-历史资产-价值,对应的训练样本集合可以表示为s
i={(C
i,P
i,V
i)|(C,P,V)∈S},i=1,2,......,n,在此不做限定。然后,根据所述资产价值训练样本对所述图卷积网络模型进行训练,如此一来,训练完成的所述图卷积网络模型可以用于资产对用户的价值进行相应的评估。
需要说明的是,对于在所述资产知识图谱中所述用户所处的节点与资产所处的节点直接连接的情况,可以是根据交互事件与资产价值之间的对应关系来确定资产对用户的价值,比如,对于资产为专利的情况,可以采用“公 司对专利的申请”、“公司之间专利的交易情况”和“公司之间专利的涉诉情况”这三个指标来评估专利价值,预先设定交互事件为“申请”时的资产价值为“8”分,交互事件为“购买”时的资产价值为“10”分,交互事件为“涉诉且胜诉”时的资产价值为“10”。
在本公开实施例中,为了实现对所述历史资产相较于所述用户的价值的计算,如图8所示,在步骤S601:确定与所述用户存在交互事件的历史资产集合之后,所述方法还包括:
S701:确定所述资产知识图谱中,所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度;
S702:根据所述交互事件以及所述路径长度,确定所述历史资产集合中每个所述历史资产对所述用户的价值,其中,所述路径长度越长,表明相应的历史资产对所述用户的价值越低。
在具体实施过程中,步骤S701至步骤S702的具体实现过程如下:
在确定与所述用户存在交互事件的历史资产集合之后,确定所述资产知识图谱中,所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度。需要说明的是,所述用户在所述资产知识图谱中对应一个实体,所述历史资产在所述资产知识图谱中对应一个实体。结合图9所示的所述资产知识图谱中的部分连接示意图来说,“公司1”可以视为用户,“专利1”、“专利2”、“专利3”以及“专利4”可以视为历史资产,“公司1”到“专利1”的路径长度为2,“公司1”到“专利3”的路径长度为5,“公司1”到“专利2”和“专利4”的路径长度为4。然后,根据路径长度与资产价值之间的对应关系,确定所述历史资产集合中每个所述历史资产对所述用户的价值,其中,所述路径长度越长,表明相应的历史资产对所述用户的价值越低,仍以图9所示为例,随着专利到公司路径长度的增加,该专利对公司的价值是递减的。特别是对于在所述资产知识图谱中所述用户所处的节点与资产所处的节点间接连接的情况,可以根据所述交互事件以及所述路径长度,来确定资产对所述用户的价值。
在本公开实施例中,步骤S702:根据所述交互事件以及所述路径长度与资产价值之间的对应关系,确定所述历史资产集合中每个所述历史资产对所述用户的价值,包括:
采用以下公式确定所述历史资产集合中每个所述历史资产对所述用户的价值:
V=A
l-1basevalue
其中,l表示所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度,l为正整数,basevalue表示基础价值,A为小于1的常数。
在具体实施过程中,若在所述资产知识图谱中,所述用户所处的节点与所述历史资产集合中任一目标历史资产所处的节点直接连接,相应地,所述用户所处的节点到所述历史资产集合中任一目标历史资产所处的节点之间的路径长度为1,此时,所述目标历史资产对所述用户的价值可以是V=basevalue,对于basevalue的具体数值可以根据所述用户与所述目标历史资产之间的交互事件来确定,比如,可以将交互事件为“申请”时的资产价值为“8”分作为基础价值basevalue,当然,还可以根据实际应用需要来设置basevalue,在此不做限定。此外,对于A的具体数值可以是0.8,还可以是0.6,在此不做限定。
在本公开实施例中,如图10所示,步骤S103:将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值,包括:
S801:确定所述资产集合中任一目标资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标资产的联系越紧密,所述相邻三元组包括至少一个三元组;
S802:以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;
S803:通过所述图卷积网络模型的拼接层将所述目标资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;
S804:将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值。
在具体实施过程中,所述图卷积网络模型可以是基于ripplenet网络的模型,结合图11所示的模型结构,对步骤S801至步骤S804的具体实现过程进行解释说明:
首先,确定所述资产集合中任一目标资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标资产的联系越紧密,所述相邻三元组包括至少一个三元组。以所述目标资产作为头实体h,并从头实体h出发搜索关系r和尾实体t,这样一个完整的三元组称为一个hop。结合图12所示的hop示意图来说,资产1的相邻三元组(即初始hop三元组)包括四个三元组,分别为资产1-关系1-实体1、资产1-关系2-实体2、资产1-关系3-实体3、资产1-属性1-属性值1,当下一个hop三元组开始时,将前一个hop三元组的尾实体作为下一个hop三元组的头实体,不断计算,实现下文提及的“传播”。
可以采用以下公式计算所述目标资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性:
p
i=softmax(v
TR
ih
i)
其中,v表示所述目标资产的嵌入向量,R和h分别表示相邻三元组中的关系和头实体的嵌入向量表示。
然后,以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述图卷积网络模型的n次传播之后,获得所述用户的最终嵌入向量,n为正整数。比如,以p
i为权重向尾实体t进行传播,其中,第一个hop三元组的输出为:
重复n-hop过程,并将所有的hop三元组的输出相加作为响应,得到所述用户的最终嵌入向量:
然后,通过所述图卷积网络模型的拼接层将所述目标资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量,所述拼接后的向量可以表示为:concat(C,v);所述拼接后的向量融合了所述目标资产的嵌入向量以及所述用户的最终嵌入向量,保证了所述目标资产和所述用户之间的信息充分融合,保证了所述图卷积网络模型对所述目标资产对所述用户的价值的针对性评估。
然后,将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值,从而通过所述图卷积网络模型实现对所述目标资产的价值的评估。经所述全连接层之后,确定了所述拼接后的向量与所述目标资产对所述用户的价值之间的映射关系,基于该映射关系可以精确地确定所述目标资产对所述用户的价值,从而提高了对用户进行资产价值评估的精确度。
在本公开实施例中,为了确定所述目标资产对所述用户的价值,步骤S804:将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值,包括:
通过所述图卷积网络模型的全连接层采用以下公式获得所述目标资产对所述用户的价值:
y=σ(concat(C,v)·ω+b)
其中,σ表示sigmoid激活函数,C表示所述用户的最终嵌入向量,v表示所述目标资产的嵌入向量,w表示权重,b表示偏置。
在本公开实施例中,由于资产的嵌入向量是基于资产与资产自身属性之间的关系,以及所述资产与其它实体之间的关系所确定的,从而提高了对用户进行资产价值评估的精确度。
基于同一公开构思,如图13所示,本公开实施例还提供了一种模型训练方法,所述模型训练方法可以是应用于资产价值评估系统,其中,所述模型训练方法包括:
S901:获取用户的历史资产交互信息;
S902:根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;
S903:对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到的,所述属性用于表征所述历史资产的固有参数;
S904:将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
在具体实施过程中,步骤S901至步骤S904的具体实现过程如下:
首先,可以是通过收集所述用户在自资产价值评估系统上的历史交互记录,根据该历史交互记录确定所述用户的历史资产交互信息,然后,根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,所述历史资产集合包括至少一个历史资产,其中,所述交互事件可以是申请、购买、涉诉、转让等,然后,对所述历史资产集合中的每个所述历史资产做嵌入表示,从而确定每个所述历史资产的资产嵌入向量,由于每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到的,所述属性用于表征所述历史资产的固有参数,所述历史资产的资产嵌入向量不仅融合了资产自身对应的实体,还融合了实体与实体自身的属性之间的关系。然后,将每个所述历史资产对应的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。由于输入至所述待训练的图卷积网络模型中的资产嵌入向量是基于每个所述历史 资产和属性的关系训练得到的,所述历史资产的资产嵌入向量融合了所述历史资产与所述历史资产自身的属性之间的关系,当用所述历史资产的资产嵌入向量对所述待训练的图卷积网络模型进行训练时,可以更精确地确定所述历史资产的相关信息,从而提高了对用于资产价值评估的模型的训练精确度。
在本公开实施例中,如图14所示,步骤S903中:所述基于每个所述历史资产和属性的关系的训练过程,包括:
S1001:确定每个所述历史资产中的格式为实体-属性-属性值的第一类三元组;
S1002:初始化所述第一类三元组的嵌入向量并进行训练,得到所述历史资产的资产嵌入向量。
在具体实施过程中,步骤S1001至步骤S1003的具体实现过程如下:
首先,确定每个所述历史资产中的格式为实体-属性-属性值的第一类三元组,所述第一类三元组的格式可以表示为(h,a,a_value),其中,h表示所述第一类三元组中的实体,a表示所述第一类三元组中的属性,a_value表示所述第一类三元组中的属性值。比如,对于三元组(专利1,法律状态,公开),“专利1”表示实体,“法律状态”表示属性,“公开”表示属性值;再比如,对于三元组(专利1,权利要求数,30),“专利1”表示实体,“权利要求数”表示属性,“30”表示“属性值”;再比如,对于三元组(专利1,涉诉次数,5),“专利1”表示实体,“涉诉次数”表示属性,“5”表示“属性值”。当然,还可以根据实际应用需要构造所述历史资产的所述第一类三元组,在此不再一一举例说明。在具体实施过程中,可以基于对资产价值影响较大的属性来构造所述第一类三元组,比如,专利的“法律状态”、“权利要求数”、“涉诉次数”往往对专利价值的影响程度较大,通过构造包括这些属性的第一类三元组,充分考虑了所述历史资产自身的属性所对应的属性值,从而能够精确地确定相应专利的价值,基于相同的实现原理,通过所述第一类三元组的构造,可以提高对用于资产价值评估的模型进行训练的精确度。
然后,初始化所述第一类三元组的嵌入向量并进行训练,得到所述历史资产 的资产嵌入向量。由于所构造的所述第一类三元组充分考虑了实体自身的属性所对应的属性值,如此一来,初始化所述第一类三元组的嵌入向量并进行训练之后,所得到的所述历史资产的资产嵌入向量充分考虑了所述历史资产与所述历史资产自身固有属性之间的关系,这样的话,基于所述历史资产的资产嵌入向量对所述待训练的图卷积网络模型进行训练,提高了对用于资产价值评估的模型进行训练的精确度。
在本公开实施例中,如图15所示,步骤S903中:所述基于每个所述历史资产和属性的关系的训练过程,还包括:
S1101:确定每个所述历史资产中的格式为头实体-关系-尾实体的第二类三元组;
S1102:通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。需要说明的是,通过所述实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量的具体实现过程,可以参照前述资产价值评估方法中的相关部分的描述,在此不再赘述。
在本公开实施例中,如图16所示,在步骤S902:根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合之后,所述方法还包括:
S1201:根据所述交互事件,确定每个所述历史资产对所述用户的价值;
S1202:根据每个所述历史资产对所述用户的价值,构建对所述待训练的图卷积网络模型进行训练的资产价值训练样本。
对于步骤S1201至步骤S1202中,确定每个所述历史资产对所述用户的价值,以及构建资产价值训练样本的具体过程,可以参照前述资产价值评估方法中的相关部分的描述,在此不再赘述。
在本公开实施例中,如图17所示,步骤S904:将每个所述历史资产对应的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型,包括:
S1301:根据预先构建的资产知识图谱,确定所述资产价值训练样本中与所述用户存在交互事件的各个所述历史资产;
S1302:确定各个所述历史资产中任一目标历史资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标历史资产的联系越紧密,所述相邻三元组包括至少一个三元组;
S1303:以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述待训练的图卷积网络模型的h次传播之后,获得的所述用户的最终嵌入向量,h为正整数;
S1304:通过所述待训练的图卷积网络模型的拼接层将所述目标历史资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;
S1305:将所述拼接后的向量输入所述待训练的图卷积网络模型中的全连接层,经所述全连接层输出所述目标历史资产对所述用户的预测资产价值;
S1306:根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值;
S1307:利用所述损失值对所述待训练的图卷积网络模型进行参数更新,获得所述图卷积网络模型。
在具体实施过程中,步骤S1301至S1307的具体实现过程如下:
首先,根据预先构建的资产知识图谱,确定所述资产价值训练样本中与所述用户存在交互事件的各个所述历史资产,对于所述资产知识图谱的构建过程可以参照前述资产价值评估方法中的相关部分的描述,在此不再赘述;然后,确定各个所述历史资产中任一目标历史资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标历史资产的联系越紧密,所述相邻三元组包括至少一个三元组;以所述目标历史资产作为头实体h,并从头实体h出发搜索关系r和尾实体t,采用以下公式计算所述目标历史资产的嵌入向量与相邻三元 组中的头实体在关系空间的相似性:p
i=softmax(v
TR
ih
i)
其中,v表示所述目标历史资产的嵌入向量,R和h分别表示相邻三元组中的关系和头实体的嵌入向量表示。
经所述待训练的图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;所述用户的最终嵌入向量为:
然后,通过所述待训练的图卷积网络模型的拼接层将所述目标历史资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量concat(C,v);然后,将所述拼接后的向量输入所述待训练的图卷积网络模型中的全连接层,经所述全连接层输出所述目标历史资产对所述用户的预测资产价值:y=σ(concat(C,v)·ω+b)
其中,σ表示sigmoid激活函数,C表示所述用户的最终嵌入向量,v表示所述目标历史资产的嵌入向量,w表示权重,b表示偏置。
然后,根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值;进一步地,利用所述损失值对所述待训练的图卷积网络模型进行参数更新,获得所述图卷积网络模型。
在本公开实施例中,步骤S1206:根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值,包括:
采用以下公式计算损失值:
其中,output表示所述预测资产价值,value表示所述预设价值,λ1和λ2表示超参数,r表示关系嵌入向量,R表示关系矩阵,I表示三元组好坏, E表示所述资产知识图谱中的实体矩阵,V表示所述待训练的图卷积网络模型的参数。
在前述对所述待训练的图卷积网络模型进行训练,获得训练完成的所述图卷积网络模型之后,当输入针对所述用户的资产价值查询信息之后,可以将对应的资产集合输入到的所述图卷积网络模型中,返回资产集合中每个资产的价值,从而实现了对不同用户的个性化价值评估。
需要说明的是,若当前存在所述用户的历史资产交互信息,可以通过在线学习的方式实现对所述用户的个性化资产价值评估,也就是说,需要训练一个基于对应用户的资产价值评估模型,进而保证了对相应用户的资产价值评估的精确度。若当前不存在所述用户的历史资产交互信息,即所述用户为新用户,比如,可以是首次申请专利的公司,还可以是首次购买手机的用户,可以记录相应的交互事件,当积累一定的历史资产交互信息之后,再调用在线学习的方式,重新训练所述待训练的图卷积网络模型的参数,后续便可以对所述用户的相关资产进行资产价值评估,进而提高了资产价值评估的精确度。
本公开实施例提供的模型训练方法解决问题的原理和前述的资产价值评估方法相似,相关实现过程可以参照前述资产价值评估方法中的相关部分的描述,重复之处不再赘述。
基于同一公开构思,本公开实施例还提供了一种资产价值评估装置,其中,包括:
第一存储器和第一处理器;
其中,所述第一存储器用于存储计算机程序;
所述第一处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:
获取输入的针对用户的资产价值查询信息;
在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;
对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;
将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
基于同一公开构思,本公开实施例还提供了一种模型训练装置,其中,包括:
第二存储器和第二处理器;
其中,所述第二存储器用于存储计算机程序;
所述第二处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:
获取用户的历史资产交互信息;
根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;
对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到,所述属性用于表征所述历史资产的固有参数;
将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
基于同一公开构思,本公开实施例还提供了一种计算机可读存储介质,其中:
所述可读存储介质存储有计算机指令,当计算机指令在计算机上运行时,使得计算机执行如上面任一项所述的资产价值评估方法或者如上面任一项所述的模型训练方法。
本领域内的技术人员应明白,本公开的实施例可提供为方法、系统、或计算机程序产品。因此,本公开可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且,本公开可采用在一个或多个 其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本公开是参照根据本公开的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器,使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中,使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品,该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上,使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理,从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
尽管已描述了本公开的优选实施例,但本领域内的技术人员一旦得知了基本创造性概念,则可对这些实施例作出另外的变更和修改。所以,所附权利要求意欲解释为包括优选实施例以及落入本公开范围的所有变更和修改。
显然,本领域的技术人员可以对本申请进行各种改动和变型而不脱离本申请的精神和范围。这样,倘若本申请的这些修改和变型属于本申请权利要求及其等同技术的范围之内,则本申请也意图包含这些改动和变型在内。
Claims (22)
- 一种资产价值评估方法,其中,包括:获取输入的针对用户的资产价值查询信息;在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
- 如权利要求1所述的方法,其中,所述基于每个所述资产和属性的关系的训练过程,包括:确定每个所述资产中的格式为实体-属性-属性值的第一类三元组;初始化所述第一类三元组的嵌入向量并进行训练,得到所述资产嵌入向量。
- 如权利要求2所述的方法,其中,所述初始化所述第一类三元组的嵌入向量,包括:通过Bi-LSTM网络对所述第一类三元组进行编码,得到所述第一类三元组的初始化嵌入向量。
- 如权利要求3所述的方法,其中,所述基于每个所述资产和属性的关系的训练过程,还包括:确定每个所述资产中的格式为头实体-关系-尾实体的第二类三元组;通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。
- 如权利要求4所述的方法,其中,所述通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组 和所述第二类三元组的嵌入向量,包括:通过所述实体和属性联合模型中的第一图嵌入Translate模型基于所述第一类三元组的初始化嵌入向量对所述第一类三元组做嵌入式训练;通过所述实体和属性联合模型中的第二图嵌入Translate模型对所述第二类三元组做嵌入式训练;基于所述第一图嵌入Translate模型的损失函数与所述第二图嵌入Translate模型的损失函数进行联合损失训练,得到所述第一类三元组和所述第二类三元组的嵌入向量。
- 如权利要求1所述的方法,其中,在所述获取输入的针对用户的资产价值查询信息之前,所述方法还包括:构建资产知识图谱,其中,所述资产知识图谱包括多个格式为实体-属性-属性值的第一类三元组和多个格式为头实体-关系-尾实体的第二类三元组。
- 如权利要求6所述的方法,其中,在所述获取输入的针对用户的资产价值查询信息之前,所述方法还包括:根据所述资产知识图谱,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;根据所述历史资产集合,构建资产价值训练样本;根据所述资产价值训练样本对所述图卷积网络模型进行训练,其中,所述资产价值训练样本中的资产对所述用户的价值是根据所述交互事件所确定的。
- 如权利要求7所述的方法,其中,所述根据所述交互事件确定所述资产价值训练样本中的资产对所述用户的价值,还包括:确定所述资产知识图谱中,所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度;根据所述交互事件以及所述路径长度,确定所述历史资产集合中每个所述历史资产对所述用户的价值,其中,所述路径长度越长,表明相应的历史资产对所述用户的价值越低。
- 如权利要求8所述的方法,其中,所述根据所述交互事件以及所述路径长度,确定所述历史资产集合中每个所述历史资产对所述用户的价值,包括:采用以下公式确定所述历史资产集合中每个所述历史资产对所述用户的价值:V=A l-1basevalue其中,l表示所述用户所处的节点到所述历史资产集合中每个所述历史资产所处的节点之间的路径长度,basevalue表示所述交互事件对应的基础价值,A为小于1的常数。
- 如权利要求1所述的方法,其中,所述将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值,包括:确定所述资产集合中任一目标资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标资产的联系越紧密,所述相邻三元组包括至少一个三元组;以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;通过所述图卷积网络模型的拼接层将所述目标资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值。
- 如权利要求10所述的方法,其中,所述将所述拼接后的向量输入所述图卷积网络模型中的全连接层,经所述全连接层输出所述目标资产对所述用户的价值,包括:通过所述图卷积网络模型的全连接层采用以下公式获得所述目标资产对所述用户的价值:y=σ(concat(C,v)·ω+b)其中,σ表示sigmoid激活函数,C表示所述用户的最终嵌入向量,v表 示所述目标资产的嵌入向量,w表示权重,b表示偏置。
- 如权利要求1所述的方法,其中,所述获取输入的针对用户的资产价值查询信息,包括:获取输入的用于表征所述用户的用户信息,以及待评估资产集合,以使所述图卷积网络模型输出所述待评估资产集合中每个待评估资产对所述用户的价值。
- 如权利要求1所述的方法,其中,所述获取输入的针对用户的资产价值查询信息,包括:获取输入的用于表征所述用户的用户信息,以使所述图卷积网络模型输出对所述用户的价值满足预设价值的至少一个资产。
- 一种模型训练方法,其中,包括:获取用户的历史资产交互信息;根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到的,所述属性用于表征所述历史资产的固有参数;将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
- 如权利要求14所述的方法,其中,所述基于每个所述历史资产和属性的关系的训练过程,包括:确定每个所述历史资产中的格式为实体-属性-属性值的第一类三元组;初始化所述第一类三元组的嵌入向量并进行训练,得到所述历史资产的资产嵌入向量。
- 如权利要求15所述的方法,其中,所述基于每个所述历史资产的实体和属性的训练过程,还包括:确定每个所述历史资产中的格式为头实体-关系-尾实体的第二类三元组;通过实体和属性联合模型对所述第一类三元组和所述第二类三元组做嵌入式训练,确定所述第一类三元组和所述第二类三元组的嵌入向量。
- 如权利要求14所述的方法,其中,在所述根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合之后,所述方法还包括:根据所述交互事件,确定每个所述历史资产对所述用户的价值;根据每个所述历史资产对所述用户的价值,构建对所述待训练的图卷积网络模型进行训练的资产价值训练样本。
- 如权利要求14所述的方法,其中,所述将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型,包括:根据预先构建的资产知识图谱,确定所述资产价值训练样本中与所述用户存在交互事件的各个所述历史资产;确定各个所述历史资产中任一目标历史资产的嵌入向量与相邻三元组中的头实体在关系空间的相似性,其中,所述相似性越高,表明所述相邻三元组中的头实体与所述目标历史资产的联系越紧密,所述相邻三元组包括至少一个三元组;以所述相似性为权重向所述相邻三元组的尾实体进行传播,经所述待训练的图卷积网络模型的n次传播之后,获得的所述用户的最终嵌入向量,n为正整数;通过所述待训练的图卷积网络模型的拼接层将所述目标历史资产的嵌入向量与所述用户的最终嵌入向量进行拼接,获得拼接后的向量;将所述拼接后的向量输入所述待训练的图卷积网络模型中的全连接层,经所述全连接层输出所述目标历史资产对所述用户的预测资产价值;根据所述预测资产价值与所述目标历史资产对所述用户的预设价值计算损失值;利用所述损失值对所述待训练的图卷积网络模型进行参数更新,获得所述图卷积网络模型。
- 一种资产价值评估装置,其中,包括:第一存储器和第一处理器;其中,所述第一存储器用于存储计算机程序;所述第一处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:获取输入的针对用户的资产价值查询信息;在确定存在所述用户的历史资产交互信息时,确定采用所述资产价值查询信息查询得到的资产集合,其中,所述资产集合包括至少一个资产;对每个所述资产做嵌入表示,确定每个所述资产的资产嵌入向量,其中,所述资产嵌入向量是基于每个所述资产和属性的关系训练得到的,所述属性用于表征所述资产的固有参数;将每个所述资产的资产嵌入向量输入图卷积网络模型中,得到每个所述资产对所述用户的价值。
- 一种模型训练装置,其中,包括:第二存储器和第二处理器;其中,所述第二存储器用于存储计算机程序;所述第二处理器用于执行所述第一存储器中的计算机程序以实现包括如下步骤:获取用户的历史资产交互信息;根据所述历史资产交互信息,确定与所述用户存在交互事件的历史资产集合,其中,所述历史资产集合包括至少一个历史资产;对每个所述历史资产做嵌入表示,确定每个所述历史资产的资产嵌入向量,其中,每个所述历史资产的资产嵌入向量是基于每个所述历史资产和属性的关系训练得到,所述属性用于表征所述历史资产的固有参数;将每个所述历史资产的资产嵌入向量输入待训练的图卷积网络模型,对所述待训练的图卷积网络模型进行训练,获得图卷积网络模型。
- 一种计算机可读存储介质,其中:所述可读存储介质存储有计算机指令,当计算机指令在计算机上运行时,使得计算机执行如权利要求1-13任一权项所述的资产价值评估方法或者如权利要求14-19任一权项所述的模型训练方法。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202180002765.7A CN116194911A (zh) | 2021-09-29 | 2021-09-29 | 资产价值评估方法、模型训练方法、装置及可读存储介质 |
PCT/CN2021/121958 WO2023050232A1 (zh) | 2021-09-29 | 2021-09-29 | 资产价值评估方法、模型训练方法、装置及可读存储介质 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2021/121958 WO2023050232A1 (zh) | 2021-09-29 | 2021-09-29 | 资产价值评估方法、模型训练方法、装置及可读存储介质 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023050232A1 true WO2023050232A1 (zh) | 2023-04-06 |
Family
ID=85781115
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/121958 WO2023050232A1 (zh) | 2021-09-29 | 2021-09-29 | 资产价值评估方法、模型训练方法、装置及可读存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN116194911A (zh) |
WO (1) | WO2023050232A1 (zh) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116596563A (zh) * | 2023-07-17 | 2023-08-15 | 北京大学 | 一种基于多因素模型的数字资产估值系统 |
CN118133963A (zh) * | 2024-05-07 | 2024-06-04 | 广东南方电信规划咨询设计院有限公司 | 一种知识子图构建方法及装置 |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101030269A (zh) * | 2006-03-03 | 2007-09-05 | 鸿富锦精密工业(深圳)有限公司 | 专利价值评估系统及方法 |
CN110457441A (zh) * | 2019-08-09 | 2019-11-15 | 佛山市木记信息技术有限公司 | 一种专利交易系统 |
CN111160783A (zh) * | 2019-12-30 | 2020-05-15 | 北京阿尔山区块链联盟科技有限公司 | 数字资产价值的评价方法、系统及电子设备 |
CN111159426A (zh) * | 2019-12-30 | 2020-05-15 | 武汉理工大学 | 一种基于图卷积神经网络的产业图谱融合方法 |
-
2021
- 2021-09-29 WO PCT/CN2021/121958 patent/WO2023050232A1/zh unknown
- 2021-09-29 CN CN202180002765.7A patent/CN116194911A/zh active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101030269A (zh) * | 2006-03-03 | 2007-09-05 | 鸿富锦精密工业(深圳)有限公司 | 专利价值评估系统及方法 |
CN110457441A (zh) * | 2019-08-09 | 2019-11-15 | 佛山市木记信息技术有限公司 | 一种专利交易系统 |
CN111160783A (zh) * | 2019-12-30 | 2020-05-15 | 北京阿尔山区块链联盟科技有限公司 | 数字资产价值的评价方法、系统及电子设备 |
CN111159426A (zh) * | 2019-12-30 | 2020-05-15 | 武汉理工大学 | 一种基于图卷积神经网络的产业图谱融合方法 |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116596563A (zh) * | 2023-07-17 | 2023-08-15 | 北京大学 | 一种基于多因素模型的数字资产估值系统 |
CN116596563B (zh) * | 2023-07-17 | 2023-09-12 | 北京大学 | 一种基于多因素模型的数字资产估值系统 |
CN118133963A (zh) * | 2024-05-07 | 2024-06-04 | 广东南方电信规划咨询设计院有限公司 | 一种知识子图构建方法及装置 |
Also Published As
Publication number | Publication date |
---|---|
CN116194911A (zh) | 2023-05-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
TWI598755B (zh) | 資料分析系統、資料分析方法、內儲資料分析程式的電腦程式產品及內儲資料分析程式的記錄媒體 | |
JP6261547B2 (ja) | 判定装置、判定方法及び判定プログラム | |
WO2023050232A1 (zh) | 资产价值评估方法、模型训练方法、装置及可读存储介质 | |
CN111061946A (zh) | 场景化内容推荐方法、装置、电子设备及存储介质 | |
WO2022179384A1 (zh) | 一种社交群体的划分方法、划分系统及相关装置 | |
JP2012058972A (ja) | 評価予測装置、評価予測方法、及びプログラム | |
US20240303507A1 (en) | Method and device for recommending goods, method and device for training goods knowledge graph, and method and device for training model | |
AU2013251195A1 (en) | Program, apparatus, and method for information processing | |
JP7229148B2 (ja) | プライバシーを保護しながら分散された顧客データ上で機械学習すること | |
JP6522050B2 (ja) | 判定装置、学習装置、判定方法及び判定プログラム | |
Atiyah et al. | Understanding Metaverse Adoption Strategy from Perspective of Social Presence and Support Theories: The Moderating Role of Privacy Risks | |
Toubøl et al. | Mapping the social class structure: From occupational mobility to social class categories using network analysis | |
KR20200047006A (ko) | 머신 러닝 기반의 근사모델 구축 방법 및 시스템 | |
US20210390263A1 (en) | System and method for automated decision making | |
CN113656699B (zh) | 用户特征向量确定方法、相关设备及介质 | |
CN115358809A (zh) | 一种基于图对比学习的多意图推荐方法及装置 | |
Praet et al. | Predictive modeling to study lifestyle politics with Facebook likes | |
KR102387829B1 (ko) | 법률 컨텐츠 제공을 위한 클라이언트 단말과 프로바이더 단말 간의 매칭 방법 | |
US20140188657A1 (en) | Establishing Customer Attributes | |
CN113343113A (zh) | 基于图卷积网络进行知识蒸馏的冷启动实体推荐方法 | |
Grincheva | Beyond the scorecard diplomacy: From soft power rankings to critical inductive geography | |
US11544327B2 (en) | Method and system for streamlined auditing | |
Cui et al. | A probabilistic-based extendable quantitative evaluation method for 3D printing service in cloud manufacturing | |
Bald et al. | spatialMaxent: Adapting species distribution modeling to spatial data | |
Hinz et al. | Integrating Reputation to Recommendation Techniques in an e-learning Environment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21958818 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 050724) |