CN110209922A - Object recommendation method, apparatus, storage medium and computer equipment - Google Patents
Object recommendation method, apparatus, storage medium and computer equipment Download PDFInfo
- Publication number
- CN110209922A CN110209922A CN201810601391.0A CN201810601391A CN110209922A CN 110209922 A CN110209922 A CN 110209922A CN 201810601391 A CN201810601391 A CN 201810601391A CN 110209922 A CN110209922 A CN 110209922A
- Authority
- CN
- China
- Prior art keywords
- vector
- attribute
- target
- feature vector
- attribute information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 95
- 239000013598 vector Substances 0.000 claims abstract description 208
- 230000006399 behavior Effects 0.000 claims description 21
- 230000015654 memory Effects 0.000 claims description 15
- 238000013528 artificial neural network Methods 0.000 claims description 12
- 238000004422 calculation algorithm Methods 0.000 claims description 11
- 238000004891 communication Methods 0.000 claims description 11
- 230000000306 recurrent effect Effects 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 7
- 238000012216 screening Methods 0.000 claims description 7
- 238000004590 computer program Methods 0.000 claims description 6
- 238000012512 characterization method Methods 0.000 claims description 5
- 230000004927 fusion Effects 0.000 claims description 5
- 241000406668 Loxodonta cyclotis Species 0.000 claims description 2
- 230000007774 longterm Effects 0.000 claims description 2
- 230000005055 memory storage Effects 0.000 claims description 2
- 238000012549 training Methods 0.000 abstract description 39
- 230000008569 process Effects 0.000 abstract description 17
- 230000014509 gene expression Effects 0.000 abstract description 13
- 238000010586 diagram Methods 0.000 description 21
- 239000000047 product Substances 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 4
- 239000011159 matrix material Substances 0.000 description 4
- 230000003542 behavioural effect Effects 0.000 description 3
- 239000000203 mixture Substances 0.000 description 3
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 230000005484 gravity Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 238000005267 amalgamation Methods 0.000 description 1
- 239000013065 commercial product Substances 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002018 overexpression Effects 0.000 description 1
- 238000002360 preparation method Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012163 sequencing technique Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000011895 specific detection Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The embodiment of the present invention provides a kind of object recommendation method, apparatus, storage medium and computer equipment, a variety of attribute informations that the present embodiment passes through acquisition target object, the feature vector for constructing the target object obtains predicted characteristics vector according to the operation order input prediction module of each target object.As it can be seen that the present embodiment enriches the expression content of feature vector, and considers the operation order of each target object in model calculating process, the accuracy of prediction result is improved;Also, due to one attribute information of multiple object-sharings of application platform output, the number of parameters of trained prediction model is greatly reduced, reduces model training difficulty, and the application scenarios of substantial amounts can be suitable for.
Description
Technical field
The present invention relates to technical field of data processing, and in particular to a kind of object recommendation method, apparatus, storage medium and meter
Calculate machine equipment.
Background technique
In recent years, popularizing for internet brings a large amount of information to user, while meeting user to information requirement,
Cause user that can not quickly find the information of useful part because information content is excessive, reduces efficiency of information.In this regard, new
The fields such as news, commercial affairs, amusement, proposition analyze behavioral data of the user on respective application platform using recommender system,
Predict the hobby of user, and filter out accordingly user's future may interested information be pushed to user, help user quickly and
Accurate selection information needed.
Wherein, in existing recommender system, it is normally based on Recognition with Recurrent Neural Network, samples multiple in application platform are used
The user behavior data at family is trained, and obtains corresponding prediction model, for predicting active user to pair in application platform
As interested probability.During model training, usually for each object in application platform, corresponding word is distributed
Vector realizes the training of prediction model as training data.
However, the quantity of the practical object having of application platform is often very big, this will lead to the parameter of prediction model
The training difficulty of prediction model has been significantly greatly increased in increasing number, and since a term vector tends not to accurately indicate in object
Hold, this will will affect the accuracy for the prediction result that prediction model actually obtains, may so as to cause the object pushed to user
It is not the interested object of user.
It can be seen that how to improve the accuracy of the recommended pushed to user, become technical staff important research side
One of to.
Summary of the invention
In view of this, the embodiment of the present invention provides a kind of object recommendation method, apparatus, storage medium and computer equipment,
Model training difficulty is reduced, and improves the accuracy of prediction result.
To achieve the above object, the embodiment of the present invention provides the following technical solutions:
A kind of object recommendation method, which comprises
The corresponding feature vector of multiple target objects is obtained, each described eigenvector characterization respective objects object
A variety of attribute informations, and the attribute information is shared by multiple target objects;
According to the operation order of the multiple target object, by corresponding feature vector sequentially input prediction model obtain it is pre-
Survey feature vector;
Calculate the similarity between predicted characteristics vector feature vector corresponding with each candidate target;
Based on the corresponding similarity of each candidate target, at least one candidate target is filtered out as recommended.
A kind of object recommendation device, described device include:
Feature vector obtains module, for obtaining the corresponding feature vector of multiple target objects, each feature
Vector characterization respective point hits a variety of attribute informations that object has, and the attribute information is shared by multiple target objects;
Predicted characteristics vector calculation module, for the operation order according to the multiple target object, by corresponding feature
Vector sequentially inputs prediction model and obtains predicted characteristics vector;
Similarity calculation module, for calculating between predicted characteristics vector feature vector corresponding with each candidate target
Similarity;
Recommended screening module filters out at least one time for being based on the corresponding similarity of each candidate target
Select object as recommended.
A kind of storage medium is stored thereon with computer program, and the computer program is executed by processor, and realizes as above
Each step of the object recommendation method.
A kind of computer equipment, the computer equipment include:
Communication interface;
Memory, for storing the program for realizing object recommendation method as described above;
Processor, for loading and executing the program of the memory storage, described program is used for:
The corresponding feature vector of multiple target objects is obtained, each described eigenvector characterization accordingly clicks object
A variety of attribute informations, and the attribute information is shared by multiple target objects;
According to the operation order of the multiple target object, by corresponding feature vector sequentially input prediction model obtain it is pre-
Survey feature vector;
Calculate the similarity between predicted characteristics vector feature vector corresponding with each candidate target;
Based on the corresponding similarity of each candidate target, at least one candidate target is filtered out as recommended.
Based on the above-mentioned technical proposal, a kind of object recommendation method, apparatus, storage medium and meter provided in an embodiment of the present invention
Machine equipment is calculated, the present embodiment constructs the feature vector of the target object by a variety of attribute informations of acquisition target object, according to
The operation order input prediction model of each target object obtains the predicted characteristics vector of the user, it is seen then that the embodiment of the present invention is rich
The rich expression content of the feature vector of target object, and the operation for considering in model calculating process each target object is suitable
Sequence improves the accuracy of prediction result;Also, due to one attribute information of multiple object-sharings of application platform output, greatly
The number of parameters for reducing trained prediction model reduces model training difficulty, and can be suitable for the applied field of substantial amounts
Scape.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
The embodiment of invention for those of ordinary skill in the art without creative efforts, can also basis
The attached drawing of offer obtains other attached drawings.
Fig. 1 is the structural schematic diagram of object recommendation system provided in an embodiment of the present invention;
Fig. 2 is a kind of flow diagram of object recommendation method provided in an embodiment of the present invention;
Fig. 3 is the structural schematic diagram for the prediction model that the present invention realizes that object recommendation method uses;
Fig. 4 is a kind of application schematic diagram of object recommendation method provided in an embodiment of the present invention;
Fig. 5 is the flow diagram of another object recommendation method provided in an embodiment of the present invention;
Fig. 6 is the application schematic diagram of another object recommendation method provided in an embodiment of the present invention;
Fig. 7 is the application schematic diagram of another object recommendation method provided in an embodiment of the present invention;
Fig. 8 is the flow diagram of another object recommendation method provided in an embodiment of the present invention;
Fig. 9 is a kind of structural schematic diagram of object recommendation device provided in an embodiment of the present invention;
Figure 10 is the structural schematic diagram of another object recommendation device provided in an embodiment of the present invention;
Figure 11 is the structural schematic diagram of another object recommendation device provided in an embodiment of the present invention;
Figure 12 is a kind of hardware structural diagram of computer equipment provided in an embodiment of the present invention.
Specific embodiment
It was found by the inventors of the present invention that user when selecting the object that application platform exports, is somebody's turn to do toward up reference
Various factors of object, by taking shopping platform as an example, when picking commodities for users, it will usually consider classification, style of commodity etc.
Various parameters, the commodity that reselection is admired can't only see the parameter of commodity one side.Based on this, inventor is proposed more
The feature-rich for meticulously expressing object, that is, use polynary semantic meaning representation mode, accurate to improve prediction to indicate an object
Property.
Also, it in order to reduce the training difficulty of prediction model, proposes by a semantic meaning representation vector, to indicate a kind of ginseng
Number, by the semantic meaning representation vector of multiple object-sharing this kind parameters, greatly reduces the semantic table of building in this way
Up to the quantity of vector, the number of parameters of prediction model is greatly reduced, the design that the present inventor proposes is learnt through practical calculating, makes
The number of parameters for obtaining prediction model is reduced to the evolution magnitude of former quantity.As it can be seen that inventor proposes above-mentioned thought, not only increase
The accuracy of prediction result, and model training efficiency is improved, achieve the purpose that improve forecasting efficiency.
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other
Embodiment shall fall within the protection scope of the present invention.
Referring to Fig.1, a kind of system structure diagram of the object recommendation method provided to realize the present invention, the system can be with
Including application server 10 and applications client 11, in which:
Application server 10, which can be, provides the service equipment of service for user, can be the clothes to match with client
Business device, if applications client is shopping software, application server can be to provide the server of shopping service, and applications client is
News browsing software, application server can be to provide server of press service etc..The present embodiment can be with application service
Device can execute present invention object recommendation method provided below, and the recommended of its needs is pushed for online user, with auxiliary
User fast and accurately selects the object needed, or user is facilitated to understand other information relevant to current browsing object, improves
Business service efficiency.
Wherein, application server 10 can be an independent application service equipment, be also possible to by multiple server structures
At service cluster, the present embodiment is not construed as limiting the structure of the server.
Applications client 11 can be mounted in such as mobile phone, laptop, iPad, the application journey in industrial personal computer equipment
Sequence specifically can be independent application program, such as software from application shop downloading installation, be also possible to web application journey
Sequence directly initiates client by application programs such as browsers, establishes the communication link with respective application server that is, without downloading
It connects.
In the present embodiment practical application, user enters the application platform of application server by applications client, can be with
Browse the application platform output various objects, in navigation process, server can according to method provided in this embodiment to
The applications client that the user uses pushes at least one recommended, which is often the interested object of user,
Object such as similar with the object that user once operated, facilitate user quickly understand interested information or selection it is interested right
As etc., it does not need user and application platform output correspondence is checked one by one again, improve efficiency of selection.
It is to be appreciated that in system provided in this embodiment, it is not limited to application server and application listed above
Client can also include other computer equipments such as multimedia server, conversation server, and above-mentioned application server 10 is logical
It often may include application database etc., selecting system can form according to actual needs, the present embodiment is no longer described in detail one by one herein.
The system structure diagram in conjunction with shown in figure 1 above, as shown in Fig. 2, the embodiment of the invention provides a kind of objects to push away
The flow diagram of method is recommended, method provided by the embodiment can be executed by application server, it can specifically include following steps:
Step S101 obtains the attribute classification of each object of application platform output;
Wherein, the attribute classification of object generally includes a variety of, and each generic attribute can be with multiple attribute informations, the present embodiment
Facilitate the attribute information for describing each generic attribute, each generic attribute can be denoted as to the first attribute, second attribute etc., corresponding classification category
The attribute information of property can be denoted as the first attribute information, second attribute information etc., wherein each first attribute information,
Two attribute informations may each comprise the attribute information of different content, and the present embodiment will not enumerate herein, it is to be appreciated that above-mentioned
" first ", " second " are not offered as order, are only used to distinguish attribute and the corresponding attribute information of attribute of all categories of all categories.
It can be seen that the present embodiment does not use single attribute and expressed when expressing an object, but use more
First expression way, that is, multi-class attribute, to describe the content of an object, so that the expression to contents of object is more accurate, in turn
Improve forecasting accuracy.
By taking application platform is shopping platform as an example, the object in application platform can be commodity, and the present embodiment can be according to
The style of commodity, by commodity be divided into such as movement, gentlewoman, white collar different-style commodity, can also be obtained according to the type of merchandise
To the multistage category information of commodity, the style of commodity can be denoted as the first attribute by the present embodiment, and the classification of commodity is denoted as second
Attribute, correspondingly, the first attribute information can be the style information of commodity, the second attribute information can be commodity classification letter
Breath, but be not limited to description of the present embodiment to the first attribute and the second attribute, that is, the attribute classification of the object obtained is not
Both attributes that the present embodiment provides are confined to, it is to be appreciated that the attribute information of each generic attribute may include different content,
It may include sports style, Feminine Style, white collar style etc. such as style attribute;Classification attribute may include shoes, cap, lining
Shirt, skirt etc..
Wherein, since commodity tend not to show on the same interface of shopping platform, usually it is divided into different stage exhibition
Show, therefore, the classification of above-mentioned commodity can be the multistage classification that the displaying rank based on commodity can be generated, therefore, this implementation
The category information of the available commodity of example can be multistage category information composition, to determine shopping platform to the displaying feelings of each commodity
Condition determines the displaying hierarchical relationship of each commodity.
Optionally, in the present embodiment, it can be detected, be come true by the attribute to all commodity in application platform
The attribute type of each object, specific detection method are not construed as limiting in the fixed application platform, later, the available application platform packet
The all properties type of each object contained can also select at least two attributes from determining attribute classification, and it is subsequent right to complete
As the building of attribute list, it is specifically chosen which attribute type does not limit.
Step S102 generates object property list according to the attribute classification of acquisition;
In the present embodiment, with the quantity classification of the object of acquisition for two kinds, i.e. for the first attribute and the second attribute into
Row explanation, the double attributes that can use acquisition generate the form of a two-dimensional table, which can be denoted as object properties
List, all rows indicate an attribute, and all column indicate another attribute, as its every a line can represent the of object
One attribute, each column can represent the second attribute of object, that is to say, that the row headers of object property list are by the first attribute
Different attribute information indicates that column heading is indicated by the different attribute information of the second attribute, the present embodiment to the first attribute at this time and
The content that second attribute includes is without limitation.
It is still illustrated by taking above-mentioned shopping platform as an example, the first attribute is the style of commodity, the class that the second attribute is commodity
Mesh, it is possible thereby to the form that the object property list constituted can be as shown in table 1 below, but it is not limited to form shown in table 1.
Table 1
In table 1, commodity nm indicates the commodity of line n m column in object property list, is obtained by the object property list
Know, each commodity can be there are two types of attribute description, and the corresponding first attribute information content of same product of doing business is identical, same
The content of corresponding second attribute information of column commodity is identical, i.e., the same product of doing business can share the first attribute letter of same content
Breath, same row commodity can share the second attribute information of same content.
It, in practical applications, can be after determining position of the commodity in object property list, by the commodity institute based on this
It is expert at and the attribute information of column describes the contents of the commodity jointly, if commodity 24 can be a ladylike hemline, commodity 11 are
Each object in application platform can be mapped to the object properties table by pair of trainers etc., in the manner described above, the present embodiment
In lattice.
Wherein, the line number of object property list and columns can be determined according to the number of objects of each generic attribute, as above
It states in citing, in all objects for including with application platform, the corresponding largest object quantity of attribute of all categories is as two-dimensional table
Line number, using categorical measure as columns, construct two-dimensional table, that is, application platform object property list.
It is to be appreciated that the representation of the object property list for application platform, it is not limited to shown in table 1 above
Form is also possible to a two-dimensional matrix, and each object in application platform can be indicated by a two-dimensional array, the present embodiment
Its content is not detailed.Also, above-mentioned object, can also be more using other other than using double attributes representation above
Meta-attribute representation, concrete methods of realizing is similar, and the present embodiment is no longer described in detail one by one herein.
Optionally, it is still illustrated so that above-mentioned first attribute is style, the second attribute is the shopping platform of classification as an example, base
In the multistage category information of commodity, the present embodiment can indicate the term vector of category informations at different levels using tree, when need
When obtaining the feature vector of object, the corresponding term vector of the corresponding category information at different levels of the available object, then to these
Term vector carries out splicing, the corresponding term vector of total classification attribute information of the object is obtained, to carry out subsequent processing.
It similarly, can also be by for the term vector of the second attribute information of column object each in above-mentioned object property list
Splice to obtain according to manner described above, i.e., the term vector of the corresponding category information of each column object in acquisition object property list,
The corresponding term vector of each object of same row is spliced again, obtains the second shared term vector of the column object, but object category
The generating mode of the corresponding term vector of each Column Properties information in property list, it is not limited to manner described herein
It is to be appreciated that the object property list of the application platform above-mentioned for the present embodiment, lower online can be constituted, so as to
The application server of application platform realizes the prediction liked active user using it.
Step S103 obtains the historical behavior data of user;
In embodiment, the object that user exports application platform operates, and the behavioral data of generation can be used as one
Historical behavior data, store into the database of the application platform, for subsequent calls, therefore, the database of application platform
In can store the historical behavior data that a plurality of different operation time generates, the present embodiment is to each historical behavior data packet
The content contained is without limitation.
Optionally, each user can be passed through with the user identifier associated storage of relative users for each historical behavior data
Mark, to distinguish the behavioral data that Object Operations of each user in application platform generate, wherein user identifier can be user
Account, terminal iidentification etc., the present embodiment to the storage form of the content of user identifier and historical behavior data without limitation.
Step S104 analyzes the historical behavior data, obtain the user multiple target objects and corresponding operation order;
It is still illustrated by taking above-mentioned shopping platform as an example, above-mentioned historical behavior data can be user and buy commodity, generate
Historical behavior data, the generation time of the historical behavior data be user buy commodity time, therefore, the present embodiment can
To buy the time sequencing of commodity according to user, arrangement is ranked up to each historical behavior data, later, to each history row
It is analyzed for data, obtains buying commodity i.e. target object accordingly.
It can be seen that the present embodiment is by analyzing historical behavior data of the user in application platform, it can be true
Which the target object of the fixed user has, and what operation order to realize the access to each target object according to, to make this
Embodiment is in the interested object of prediction user, it can be considered that user is once suitable to the access of each object in application platform
Sequence, and then the interest transition and interest accumulation of user are accurately positioned.
Step 105, according to the object property list, the row attribute letter that each target object is located in object property list is obtained
Breath and Column Properties information;
Step S106 obtains the feature vector of respective objects object according to the row attribute information and Column Properties information;
It optionally, can be from this for the either objective object of user in conjunction with the description above for object property list
It obtains the target object in object property list to be expert at and the corresponding attribute information of column, at once attribute information and Column Properties
Information, to generate the feature vector of respective objects object using both attribute informations, specific generation method is not construed as limiting,
It is referred to the description of hereafter corresponding embodiment, but is not limited to the implementation method being set forth below.
It can be seen that the feature vector for each target object that the present embodiment obtains characterizes a variety of categories of respective objects object
Property information, that is to say, that the content of the target object of the present embodiment is described by a variety of attribute informations, and each attribute information pair
Answer a generic attribute.
It is to be appreciated that the method about the corresponding feature vector for obtaining target object, it is not limited to which the present embodiment is retouched
This implementation method using object property list stated, also can use the modes such as matrix, array and realizes, implementation method class
Seemingly, this will not be detailed here for the present embodiment.
Corresponding feature vector is sequentially input prediction model and obtained by step S107 according to the operation order of each target object
To predicted characteristics vector;
Optionally, the prediction model of the present embodiment can be based on Recognition with Recurrent Neural Network, to the spy of multiple sample object objects
Sign vector training obtains, which can wrap containing multiple shot and long term memory network (LSTM, Long Short-
Term Memory) layer, i.e., the hidden layer of Recognition with Recurrent Neural Network is constituted by multiple LSTM layers, is realized and is hidden using the principle of LSTM
The calculating of layer, the present embodiment are not described in detail here.
Wherein, LSTM is used to solve conventional recycle neural network for long-term Dependence Problem, by introducing gate list
Member, by certain study, can be acquired to handle memory/forgetting of memory unit, input degree, export the problem of degree
When open which kind of degree arrived to each.
Referring to the structural schematic diagram of prediction model shown in Fig. 3, input layer, hidden layer and output layer can be divided into, it is defeated
Enter layer and input corresponding feature vector of each moment, such as the X in Fig. 3t-1、Xt、Xt+1Deng, hidden layer may include multiple LSTM layers (such as
The structure in each box in Fig. 3), it is found that in Recognition with Recurrent Neural Network, the output of last moment adds structure as shown in Figure 3
The input at upper current time obtains the output of subsequent time by tanh activation primitive, and the addition of LSTM, so that hidden layer
Increase three doors, it is specific its by the output h of last momentt-1With the input X at current timetAs a whole, as current
The input of hidden layer, while the input of three doors is controlled, forgetting door is first passed through, obtained remnant continues forward, in addition this
Input information under a state (does tanh operation to input in the intermediate box in such as Fig. 3, codomain is controlled in [- 1,1] range
It is interior), the remaining department by being input gate treated input signal and a upper door adds up, and obtains the cell at current time
State completes the update from last moment to current time.The function of last out gate is that it is postrun hiding with tanh
Layer state information is multiplied, and obtains the output at current time.It is to be appreciated that σ is indicated in LSTM in respective doors calculating process in Fig. 3
Coefficient, the present embodiment are not construed as limiting its numerical value, and h indicates the output of corresponding moment hidden layer.
Based on foregoing description, the output of the hidden layer at current time can be defeated with last moment hidden layer by currently inputting
It constitutes out, specifically inner product, inner product can be asked with the term vector of each generic attribute respectively by the predicted characteristics vector being currently calculated
It is bigger, illustrate to get over preference to the object of respective attributes, it is possible thereby to determine the weight of both attributes, but the calculating of hidden layer
Journey is not limited to this implementation, and the present embodiment is no longer described in detail one by one herein, and for the structure of prediction model not office
It is limited to structural schematic diagram shown in Fig. 3.
As it can be seen that the expression of the hidden layer at each moment can be, according to the behavior sequence information of target object before the moment
(feature vector of such as above-mentioned each target object), and the preference expression of target object is obtained, it is to be appreciated that obtained user
Preference expression, might not correspond to existing object, can be the approximate expression of a rarefaction.
For example, user A handles its historical behavior data in the manner described above, is obtained by prediction model
The predicted characteristics vector arrived, can express that the user A next moment most wants to buy is shoes class, move wind clothes (more than
It is illustrated for the corresponding shopping platform of table 1), then, by hidden layer it can be calculated that the preference of user instantly is closest
Sport footwear.
In the present embodiment practical application, before application server carries out object prediction to user, use can be trained in advance
To realize that the prediction model of prediction calculating, the present embodiment are referred to aforesaid way, obtain the target pair of multiple sample of users
The feature vector of elephant is based on Recognition with Recurrent Neural Network algorithm using these feature vectors as training data, to these training datas into
Row constantly training, until model Complete Convergence, i.e. prediction result and actual result is essentially identical, and error in a certain range, can
With deconditioning, using finally obtained model as prediction model.The present embodiment is not described further above-mentioned model training process.
Also, in conjunction with the description above to LSTM principle, in the training process of prediction model, the present embodiment be can use
The training data of the multiple sample of users obtained carries out gradient anti-pass study to the parameter in prediction model, that is, uses boarding steps
Descent method is spent, it will be residual by gradient from functional relation according to the residual error (remnant described above) from objective result
Difference is transmitted to each layer of Recognition with Recurrent Neural Network, modifies the parameters of Recognition with Recurrent Neural Network, so that the result of output be made more to connect
Nearly legitimate reading, details are not described herein for concrete application process the present embodiment.
Wherein, during about model training, the acquisition of training data is referred to the feature vector of target object above
Acquisition modes, to obtain the feature vector of each sample target object, this will not be detailed here for the present embodiment.
Furthermore, it is necessary to illustrate, if the data volume of the object of the application platform output obtained is insufficient and limited calculated amount
In the case of, the present invention can also use other modes, be trained to above-mentioned training data, obtain the prediction mould of other structures
Type, such as Markov model, collaborative filtering, matrix decomposition scheduling algorithm realize the model training to training data, specific training
Process can determine that the present embodiment does not do be described in detail one by one herein based on the principle of each algorithm.It can be seen that for above-mentioned prediction mould
Type is not limited to model structure shown in figure 3 above, and the present embodiment is only illustrated as example.
Step S108, the corresponding feature vector of each candidate target and predicted characteristics vector that computing object attribute list includes
Between similarity;
Wherein, the present embodiment can be using each object in application platform as candidate target, in object property list
Have corresponding attribute information, can mode as described above, obtain the corresponding feature vector of each candidate target, this implementation
Details are not described herein for example.
Optionally, the target object of the present embodiment can be the click object that user carries out clicking operation, accordingly it is also possible to
The factors such as the clicking rate and the clicking rate on other application platform of the object based on application platform output, it is defeated to application platform
All objects out are screened, and multiple candidate targets are obtained, and the feature vector of each candidate target is obtained according still further to aforesaid way,
To the multiple objects how to export from application platform, the method for obtaining candidate target is not construed as limiting the present embodiment.It is to be appreciated that should
Candidate target may include the target object of non-present output, and user is not limited to click to the operation of object, for it
The operation of his mode, the foundation for screening candidate target can be adjusted accordingly, it is not limited to the screening mode of the present embodiment description.
In addition, the present embodiment to the similarity calculating method between vector without limitation, as cosine similarity calculation method,
Distance calculating method etc. can realize that the present embodiment is no longer described in detail one by one herein according to corresponding similarity calculating method principle.
Step S109 filters out at least one candidate target as recommended based on the similarity of each candidate target.
Wherein, the similarity being calculated is higher, indicate corresponding candidate object be screened for the probability of recommended it is bigger,
The present embodiment realizes that the screening technique of recommended is not construed as limiting to how based on similarity height.
Optionally, it can be ranked up according to the similarity height for each candidate target being calculated, to select similar
Recommended of the highest certain amount of candidate target as user, the i.e. sequence according to similarity from high to low are spent, is selected
Certain amount of candidate target is as recommended out, it is of course also possible to similarity threshold be preset, to select similarity
Greater than the similarity threshold candidate target as recommended etc..
In the present embodiment practical application, the recommended filtered out can be pushed into user client and be shown,
To assist user's quick-pick to go out suitable object.
To sum up, referring to schematic diagram shown in Fig. 4, the present embodiment obtains a variety of attribute informations of target object, constructs target
The predicted characteristics vector of user is calculated as the input of prediction model in the feature vector of object, not only enrich feature to
The expression content of amount, and due to one attribute information of multiple object-sharings of the present embodiment application platform output, so that constituting each
Attribute information needed for the feature vector of target object, may be reused, and greatly reduces the number of parameters of prediction model, keeps away
Exempt to occur being overexpressed the problem of serious and Sparse is cold-started, so that model training becomes to be more easier, and can be suitable for
The application scenarios of substantial amounts;Moreover, the present embodiment, which is based on Recognition with Recurrent Neural Network, carries out model training, it is also contemplated that when data
Sequence improves the accuracy of prediction result.
As an alternate embodiment of the present invention, another kind is proposed according to object property list, obtains each target object pair
The process for the feature vector answered, it is also assumed that being the concrete methods of realizing of above-mentioned steps S105 and step S106, but step
The implementation method of S105 and S106 is not limited to this implementation of the present embodiment description, the object category about application platform
The generating process of property list, is referred to the description of above-described embodiment corresponding portion, and how the present embodiment is mainly to obtaining target
The realization process of the feature vector of object is described, flow diagram as shown in Figure 5 and Figure 6, this method may include but
It is not limited to following steps:
Step S201 obtains location information of each target object in object property list;
Step S202 is based on the location information, obtains respective objects object the first term vector of the row and column
Second term vector;
For object property list shown in the above table 1, if target object is commodity 24, the commodity can be searched in object
Position in attribute list, i.e. the 2nd row the 4th column, later, the available corresponding attribute information of the 2nd row of object property list
The category of the shared row attribute information of (i.e. this row attribute information of gentlewoman's wind), the i.e. each commodity of the row of commodity 24 and the 4th column
Property information (i.e. this Column Properties information of skirt), that is, the shared Column Properties information of each commodity of 24 column of commodity.
Later, the first term vector can be generated by the row attribute information of obtained target object in the present embodiment, by dependent of dead military hero
Property information generate the second term vector, specific generating process is without limitation.By above-mentioned object property list it is found that same a line pair
As sharing a row attribute information, and then same first term vector is shared, in one Column Properties letter of object-sharing of same row
Breath, and then shares same second term vector, and therefore, the present embodiment is not needed for each object, all the of building object
One term vector and the second term vector, the step of simplifying the initialization semantic meaning representation to object each in object property list.Wherein,
The term vector of the present embodiment can be the semantic meaning representation mode of object respective attributes information, and how the present embodiment is to being believed using attribute
Breath, the method for generating corresponding term vector are not construed as limiting.
For object property list shown in the above table 1, the of movement wind (i.e. a kind of row attribute information) this product of doing business
One term vector can obtain first term vector of this attribute information line by line by carrying out semantic analysis to movement;Similarly, may be used
To obtain gentlewoman's wind (i.e. another row attribute information) corresponding first term vector of this product of doing business, shoes (i.e. a kind of Column Properties
Information) the corresponding Column Properties information of this column commodity the second term vector, skirt (i.e. another Column Properties information) this column quotient
Second term vector of the corresponding Column Properties information of product, and so on, it can be by first term vector for each commodity
Co-expressed with second term vector, and be located at first term vector of each commodity of a line it is identical, positioned at same row
Second term vector of each commodity is identical.
It is to be appreciated that the present embodiment to building object property list in, the first term vector of each row object and each column object
The concrete methods of realizing of the second term vector be not construed as limiting.
Step S203, the first term vector corresponding to same target object and the second term vector merge, and obtain each mesh
Mark the corresponding feature vector of object.
Side of the present embodiment to the fusion for how realizing corresponding first term vector of same target object and the second term vector
Method is not construed as limiting, for example directly splice the first term vector of same target object and the second term vector, will obtain to
Feature vector of the amount as user click vector, so that this feature vector be enable to characterize two generics of the target object simultaneously
Property information.
Optionally, the present embodiment can use door control unit, realize corresponding first term vector of above-mentioned each target object and
The fusion of second term vector, the door control unit are intended to handle and select under existing object expression, row attribute information and Column Properties letter
It ceases shared specific gravity in user's selection next time movement, and according to such specific gravity goes to merge both attribute informations corresponding
Term vector, so that obtaining an optimal recommended characteristics expresses vector, the present embodiment is denoted as feature vector.As it can be seen that this reality
It applies example and is merged two kinds of term vectors by door control unit, simulate the thinking characteristic in user's access Object Process, have more
Strong flexibility meets the personalized recommendation demand of different user.
In conclusion the present embodiment obtains a variety of attributes of target object when constructing the feature vector of target object
Information architecture enriches the contents of object of feature vector expression, and due to being located in object property list with a line or same row
Object be located at when clicking the feature vector of object with a line or same row due to sharing same attribute information so that calculating,
The same term vector can be shared, the number of parameters of prediction model is greatly reduced, improves forecasting efficiency and accuracy.
Optionally, the present embodiment obtains the feature vector process of each sample target object during training prediction model
In, it can also construct in the manner described above, i.e., the mode construction feature vector of shared term vector greatly reduces training pattern
Number of parameters greatly reduces the training difficulty of prediction model, moreover, because a variety of attributes for capableing of combining target object are constituted
Training data, it is contemplated that the minutia of object, such as above-mentioned style information and it is divided into classification information, so that training be made to obtain
Prediction model prediction it is more accurate and refine.
Optionally, in order to which the forecasting accuracy for improving prediction model is carrying out prediction mould on the basis of the above embodiments
In the training process of type, can also the position in the following way to each object in object property list optimize, but simultaneously
It is not limited to method provided in this embodiment, as shown in fig. 7, this method may include:
Step S301 obtains the row attribute information of practical object;
Wherein, practical object can be the above-mentioned recommended filtered out, such as the commodity of user's actual purchase.
Step S302 calculates the row attribute letter of multiple object for multiple objects of either rank in object property list
The similarity of breath and the row attribute information of practical object;
The present embodiment is not construed as limiting the similarity calculating method of the row attribute information between two objects, such as generates corresponding
Term vector after, calculate the similarity between corresponding term vector, to indicate similarity that the two objects are expert on attribute, if
Row attribute is the style attribute of commodity, which can indicate similarity of two objects in style.
Step S303, based on the similarity being calculated, according to bipartite graph matching algorithm, to each in the object property list
The position of object is adjusted.
For example, recommended is three pairs of shoes, it is denoted as q1, q2, q3 respectively, using in the training process of prediction model, if with
Family purchase is q1, obtains this three pairs of shoes in style attribute s1, and the upper similarity of s2, s3 is (0.1,0.5,0.1) respectively, (0.3,
0.1,0.2), (0.3,0.3,0.7) according to bipartite graph matching algorithm, can learn that q1 matches s2 later, and q2 matches s1, q3
When with s3, therefore q1 can be distributed to style attribute s2 by obtaining global similarity highest, q2 distributes to style attribute
S1, q3 distribute to style attribute s3.Wherein, bipartite graph matching algorithm can use Minimum-cost flow problem, and
Using the map analysis algorithm packet of google OR-tools, realize to each object's position in object property list obtained above
It optimizes, to improve the accuracy for the object prediction realized accordingly.
In conjunction with above-mentioned analysis, the present invention is by taking commercial product recommending scene as an example, should to illustrate the application of above-mentioned object recommendation method
Application platform under scene can be shopping platform, and target object can be the commodity of user's purchase, this is based on, referring to Fig. 8 institute
The flow diagram shown, user can send access request to application server by client, log in shopping platform, browsing purchase
The extensive stock of object platform output, corresponds to for application server at this time, after the access request for receiving user, can be based on
The user identifier carried in the access request obtains the historical behavior data that the user buys commodity, by analyzing it,
It can learn that the commodity that user once bought, the present embodiment can be denoted as history commodity, later, for each history quotient
Product can determine its position in the object property list that the shopping platform has constructed, thus where obtaining the history commodity
The category information of capable style information and column, and then corresponding first term vector and the second term vector are generated, by fusion
Processing, obtains the feature vector of the history commodity.
Later, the corresponding feature vector input of each history commodity can have been instructed according to the order of purchase of each history commodity
The prediction model perfected, output vector are predicted characteristics vector, can be as choosing for predicting the purchase preference of user
The standard of Recommendations is selected, the similarity between the feature vector and the predicted characteristics vector of each candidate commodity, similarity are calculated
It is higher, show that a possibility that user buys candidate's commodity is bigger, therefore, the present embodiment can choose similarity highest K
As Recommendations, the client for being sent to user is shown candidate commodity, so that user is referring to selection purchase commodity.
It wherein, is still the feature that multiple sample of users purchase commodity are obtained using aforesaid way in building prediction model
Vector, i.e. the potential applications expression of refinement modeling object, to improve the recommendation forecasting accuracy of the prediction model.And by level
Classification attribute and attributes preferred (i.e. style attribute) decoupling are opened, and corresponding semantic meaning representation are determined as independent attribute, to make pre-
The habit of user can clearly be analyzed by surveying model, additionally it is possible to based on obtained prediction classification and attributes preferred, it is known that other works
Make, such as collocation of commodity etc.,
It is a kind of structure chart of object recommendation device provided in an embodiment of the present invention referring to Fig. 9, which can be applied to
Application server can specifically include but be not limited to consisting of structure:
Feature vector obtains module 91, for obtaining the corresponding feature vector of multiple target objects;
Wherein, a variety of attribute informations that each feature vector characterization respective objects object has, and the attribute information quilt
Multiple target objects are shared.
Predicted characteristics vector calculation module 92 will be corresponding special for the operation order according to the multiple target object
Sign vector sequentially inputs prediction model and obtains predicted characteristics vector;
In the present embodiment, the prediction model can be by the corresponding feature vector of multiple sample object objects
It is trained to obtain, can specifically be realized based on modes such as Recognition with Recurrent Neural Network, collaborative filtering, matrix decompositions, the present embodiment pair
The specific training method of the prediction model is not construed as limiting.
In conjunction with above-mentioned analysis it is found that used training data contains a variety of attribute informations when carrying out model training,
So that training characteristics are more abundant, and the class object in application platform can share an attribute information, not need for each
Object sets special attribute information, reduces the quantity of model parameter, reduce model training difficulty, and it is pre- to improve model
Survey accuracy.
Similarity calculation module 93, for calculating between predicted characteristics vector feature vector corresponding with each candidate target
Similarity;
Recommended screening module 94 filters out at least one for being based on the corresponding similarity of each candidate target
Candidate target is as recommended.
Wherein, the similarity of the candidate target is higher, be screened for the probability of recommended it is bigger.
It, such as can be with it is to be appreciated that the present embodiment is not construed as limiting the method how to filter out recommended based on similarity
According to the similarity sequence from high to low of each candidate target, select certain amount of candidate target as recommended,
Alternatively, the candidate target that screening similarity reaches preset threshold is recommended etc..
Optionally, as shown in Figure 10, which can also include:
Attribute obtains module 95, for obtaining the attribute classification of each object;
Wherein, the attribute classification of object at least may include the first attribute and the second attribute.
Object property list generation module 96 generates object property list for the attribute classification according to acquisition;
Wherein, the row headers of the object property list are indicated by the different attribute information of the first attribute of each object, column
Title indicates that the specific composition method about object property list is referred to by the different attribute information of the second attribute of object
The description of above method embodiment, form as listed in Table 1, but it is not limited to form shown in table 1.
Correspondingly, described eigenvector acquisition module 91 may include:
Attribute information acquiring unit, for obtaining each target object and being located at the object category according to object property list
Row attribute information and Column Properties information in property list;;
Feature vector acquiring unit, for obtaining respective objects object according to the row attribute information and Column Properties information
Feature vector.
As another alternative embodiment, as shown in figure 11, above-mentioned object recommendation device can also include:
Data acquisition module 97, for obtaining the historical behavior data of user, which is based on user couple
What the operation that application platform exports object generated;
Data analysis module 98 obtains multiple target objects of the user, and more for analyzing the historical behavior data
The operation order of a target object;
Described eigenvector obtains module 91
Location information acquiring unit 911, for obtaining location information of each target object in the object property list;
Term vector acquiring unit 912, for be based on the location information, obtain respective objects object the first word of the row to
Second term vector of amount and column;
Optionally, if the first attribute includes the style of object, the second attribute includes the classification of the object, object properties column
Second term vector of each column object in table, can first by each object of the column category information generate term vector, then to its into
Row fusion (such as connecting method, but be not limited to this amalgamation mode) obtains, and similarly, the first term vector can also be according to this
Mode obtains, and is based on this, above-mentioned object recommendation device can also include:
First term vector obtain module, for obtain the object property list with object each in a line it is corresponding first belong to
Property information;
First term vector generation module, for generating the term vector of corresponding object by first attribute information;
First term vector splicing module obtains described for splicing to the corresponding term vector of object each in same a line
The first shared term vector of each object in the object property list row.
Second term vector obtains module, belongs to for obtaining each object of object property list same row kind corresponding second
Property information;
Second term vector generation module, for generating the term vector of corresponding object by the second attribute information;
Second term vector splicing module obtains described for splicing to the corresponding term vector of object each in same row
The second shared term vector of each object in the object property list column.
Term vector integrated unit 913, for first term vector and second term vector to same target object
It is merged, obtains the corresponding feature vector of each target object.
Optionally, term vector integrated unit 913 specifically can be used for first term vector of same target object and
Second term vector is spliced, and the feature vector of the target object is obtained;Alternatively, by door control unit, to same mesh
First term vector and second term vector for marking object are merged, and the feature vector of respective objects object is obtained, but
It is not limited to both fusion treatment modes.
Optionally, on the basis of the various embodiments described above, object recommendation device can also include:
Practical object attribute information obtains module, and for obtaining the row attribute information of practical object, the practical object is
The recommended filtered out;
Row attributes similarity computing module is calculated for multiple objects for either rank in the object property list
The similarity of the row attribute information of the row attribute information and practical object of multiple object;
Module is adjusted, for arranging according to bipartite graph matching algorithm the object properties based on the similarity being calculated
The position of each object is adjusted in table.
In the present embodiment practical application, the object property list based on generation during carrying out model training, is instructed every time
White silk terminates, and the prediction result that can be obtained based on this training is compared with actual result, based on comparative result, by object
All objects in attribute list are mapped to the maximum position of possibility, to realize the update to object property list.
Moreover, this prediction result and practical operation pair can also be based on after carrying out actual prediction using prediction model
As, continue the position of object in regulating object attribute list, it is and right so that the position of each object in the object property list
Should row attribute and corresponding Column Properties most match, i.e., object attribute information of the row and the attribute information of column can be most accurate
The contents of object is expressed, to improve the subsequent accuracy predicted based on the object property list.
In conclusion the present embodiment obtains a variety of attribute informations of target object, the feature vector of the target object is constructed,
As the input of prediction model, the predicted characteristics vector of user is calculated, not only enriches the expression content of feature vector, and
And due to one attribute information of multiple object-sharings of the present embodiment application platform output so that constitute the feature of each target object to
Attribute information needed for amount, may be reused, greatly reduce the number of parameters of prediction model, and it is serious to avoid the occurrence of overexpression
The problem of with Sparse cold start-up, so that model training becomes to be more easier, and can be suitable for the applied field of substantial amounts
Scape;It is when carrying out user's interested object properties prediction, it is also contemplated that data time sequence improves the standard of prediction result
True property.
The embodiment of the invention also provides a kind of storage mediums, are stored thereon with computer program, the computer program quilt
Processor executes, and realizes that each step of above-mentioned object recommendation method, the realization process of the object recommendation method are referred to above-mentioned
The description of embodiment of the method.
As shown in figure 12, the embodiment of the invention also provides a kind of hardware structural diagram of computer equipment, the calculating
Machine equipment can be the application server for realizing above-mentioned object recommendation method, may include communication interface 121,122 and of memory
Processor 123;
In embodiments of the present invention, communication interface 121, memory 122, processor 123 can be realized by communication bus
Mutual communication, and the communication interface 121, memory 122, processor 123 and communication bus quantity can be at least one
It is a.
Optionally, communication interface 121 can be the interface of communication module, such as the interface of gsm module;
Processor 123 may be a central processor CPU or specific integrated circuit ASIC (Application
Specific Integrated Circuit), or be arranged to implement the integrated electricity of one or more of the embodiment of the present invention
Road.
Memory 122 may include high speed RAM memory, it is also possible to further include nonvolatile memory (non-
Volatile memory), a for example, at least magnetic disk storage.
Wherein, memory 122 is stored with program, the program that processor 123 calls memory 122 to be stored, on realizing
State each step of the object recommendation method applied to computer equipment;
Optionally, which is primarily useful for:
The corresponding feature vector of multiple target objects is obtained, each feature vector characterizes a variety of of respective objects object
Attribute information, and the attribute information is shared by multiple target objects;
According to the operation order of the multiple target object, by corresponding feature vector sequentially input prediction model obtain it is pre-
Survey feature vector;
Calculate the similarity between predicted characteristics vector feature vector corresponding with each candidate target;
Based on the corresponding similarity of each candidate target, at least one candidate target is filtered out as recommended.
It is to be appreciated that executing other steps that program realizes object recommendation method about processor, it is referred to above-mentioned side
The description of method embodiment corresponding portion.
Each embodiment in this specification is described in a progressive manner, the highlights of each of the examples are with other
The difference of embodiment, the same or similar parts in each embodiment may refer to each other.For device disclosed in embodiment,
For storage medium, computer equipment, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, phase
Place is closed referring to method part illustration.
Professional further appreciates that, unit described in conjunction with the examples disclosed in the embodiments of the present disclosure
And algorithm steps, can be realized with electronic hardware, computer software, or a combination of the two, in order to clearly demonstrate hardware and
The interchangeability of software generally describes each exemplary composition and step according to function in the above description.These
Function is implemented in hardware or software actually, the specific application and design constraint depending on technical solution.Profession
Technical staff can use different methods to achieve the described function each specific application, but this realization is not answered
Think beyond the scope of this invention.
The step of method described in conjunction with the examples disclosed in this document or algorithm, can directly be held with hardware, processor
The combination of capable software module or the two is implemented.Software module can be placed in random access memory (RAM), memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, register, hard disk, moveable magnetic disc, CD-ROM or technology
In any other form of storage medium well known in field.
The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention.
Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein
General Principle can be realized in other embodiments in the case where not departing from core of the invention thought or scope.Therefore, originally
Invention is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein
Consistent widest scope.
Claims (13)
1. a kind of object recommendation method, which is characterized in that the described method includes:
The corresponding feature vector of multiple target objects is obtained, each described eigenvector characterizes a variety of of respective objects object
Attribute information, and the attribute information is shared by multiple target objects;
According to the operation order of the multiple target object, corresponding feature vector is sequentially input into prediction model and obtains prediction spy
Levy vector;
Calculate the similarity between predicted characteristics vector feature vector corresponding with each candidate target;
Based on the corresponding similarity of each candidate target, at least one candidate target is filtered out as recommended.
2. the method according to claim 1, wherein the prediction model is based on Recognition with Recurrent Neural Network to multiple samples
The corresponding feature vector of this target object, which is trained, to be obtained, and the Recognition with Recurrent Neural Network includes multiple shot and long terms
Memory network layer.
3. method according to claim 1 or 2, which is characterized in that the method also includes:
The attribute classification of each object is obtained, the attribute classification includes at least the first attribute and the second attribute;
According to the attribute classification of acquisition, object property list is generated, wherein the row headers of the object property list are by described the
The different attribute information of one attribute indicates that column heading is indicated by the different attribute information of second attribute.
4. according to the method described in claim 3, it is characterized in that, it is described obtain the corresponding feature of multiple target objects to
Amount, comprising:
According to the object property list, obtain row attribute information that each target object is located in the object property list and
Column Properties information;
According to the row attribute information and Column Properties information, the feature vector of respective objects object is obtained.
5. according to the method described in claim 3, it is characterized in that, the method also includes:
The historical behavior data of user are obtained, the historical behavior data are the operations based on user to application platform output object
It generates;
The historical behavior data are analyzed, the operation of the multiple target objects and the multiple target object of the user is obtained
Sequentially;
It is described to obtain the corresponding feature vector of multiple target objects, comprising:
Obtain location information of each target object in object property list;
Based on the location information, the second term vector of respective objects object the first term vector of the row and column is obtained;
First term vector and second term vector of same target object are merged, it is corresponding to obtain each target object
Feature vector.
6. according to the method described in claim 3, it is characterized in that, the method also includes:
The row attribute information of practical object is obtained, the practical object is the recommended filtered out;
For multiple objects of either rank in the object property list, row attribute information and the reality for calculating multiple object are right
The similarity of the row attribute information of elephant;
Based on the similarity being calculated, according to bipartite graph matching algorithm, to the position of each object in the object property list
It is adjusted.
7. method according to claim 1 or 2, which is characterized in that the similarity based on each candidate target, sieve
At least one candidate target is selected as recommended, comprising:
According to the similarity sequence from high to low of each candidate target, certain amount of candidate target is selected to push away as described in
Recommend object.
8. according to the method described in claim 5, it is characterized in that, being carried out to first term vector and second term vector
Fusion treatment obtains the corresponding feature vector of the target object, comprising:
First term vector and second term vector are spliced, the corresponding feature vector of the target object is obtained;
Alternatively,
By door control unit, first term vector and second term vector are merged, the target object pair is obtained
The feature vector answered.
9. according to the method described in claim 5, it is characterized in that, the method also includes:
Obtain the object property list with object each in a line corresponding first attribute information;
By first attribute information, the term vector of corresponding object is generated;
The corresponding term vector of object each in same a line is spliced, each object in the object property list row is obtained and shares
The first term vector.
10. according to the method described in claim 5, further include:
Obtain corresponding second attribute information of each object in the object property list same row;
By the second attribute letter information, the term vector of corresponding object is generated;
The corresponding term vector of object each in same row is spliced, each object in the object property list column is obtained and shares
The second term vector.
11. a kind of object recommendation device, which is characterized in that described device includes:
Feature vector obtains module, for obtaining the corresponding feature vector of multiple target objects, each described eigenvector
A variety of attribute informations that characterization respective objects object has, and the attribute information is shared by multiple target objects;
Predicted characteristics vector calculation module, for the operation order according to the multiple target object, by corresponding feature vector
It sequentially inputs prediction model and obtains the predicted characteristics vector of corresponding target object;
Similarity calculation module, for calculating the phase between predicted characteristics vector feature vector corresponding with each candidate target
Like degree;
It is right to filter out at least one candidate for being based on the corresponding similarity of each candidate target for recommended screening module
As recommended.
12. a kind of storage medium, is stored thereon with computer program, which is characterized in that the computer program is held by processor
Row realizes each step of the object recommendation method as described in claim 1-10 any one.
13. a kind of computer equipment, which is characterized in that the computer equipment includes:
Communication interface;
Memory, for storing the program for realizing the object recommendation method as described in claim 1-10 any one;
Processor, for loading and executing the program of the memory storage, described program is used for:
The corresponding feature vector of multiple target objects is obtained, each described eigenvector characterizes a variety of of respective objects object
Attribute information, and the attribute information is shared by multiple target objects;
According to the operation order of the multiple target object, corresponding feature vector is sequentially input into prediction model and obtains prediction spy
Levy vector;
Calculate the similarity between predicted characteristics vector feature vector corresponding with each candidate target;
Based on the corresponding similarity of each candidate target, at least one candidate target is filtered out as recommended.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810601391.0A CN110209922B (en) | 2018-06-12 | 2018-06-12 | Object recommendation method and device, storage medium and computer equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810601391.0A CN110209922B (en) | 2018-06-12 | 2018-06-12 | Object recommendation method and device, storage medium and computer equipment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110209922A true CN110209922A (en) | 2019-09-06 |
CN110209922B CN110209922B (en) | 2023-11-10 |
Family
ID=67778844
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810601391.0A Active CN110209922B (en) | 2018-06-12 | 2018-06-12 | Object recommendation method and device, storage medium and computer equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110209922B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110737834A (en) * | 2019-10-14 | 2020-01-31 | 腾讯科技(深圳)有限公司 | Business object recommendation method and device, storage medium and computer equipment |
CN110807150A (en) * | 2019-10-14 | 2020-02-18 | 腾讯科技(深圳)有限公司 | Information processing method and device, electronic equipment and computer readable storage medium |
CN110955831A (en) * | 2019-11-25 | 2020-04-03 | 北京三快在线科技有限公司 | Article recommendation method and device, computer equipment and storage medium |
CN111079015A (en) * | 2019-12-17 | 2020-04-28 | 腾讯科技(深圳)有限公司 | Recommendation method and device, computer equipment and storage medium |
CN111292168A (en) * | 2020-02-06 | 2020-06-16 | 腾讯科技(深圳)有限公司 | Data processing method, device and equipment |
CN111339425A (en) * | 2020-03-05 | 2020-06-26 | 拉扎斯网络科技(上海)有限公司 | Object marking method, device, server and storage medium |
CN111461825A (en) * | 2020-03-30 | 2020-07-28 | 京东数字科技控股有限公司 | Virtual resource generation method and device, electronic equipment and storage medium |
CN112215684A (en) * | 2020-10-30 | 2021-01-12 | 腾讯科技(深圳)有限公司 | Clustering method and device for target controllable objects |
CN112507216A (en) * | 2020-12-01 | 2021-03-16 | 北京奇艺世纪科技有限公司 | Data object recommendation method, device, equipment and storage medium |
CN112597395A (en) * | 2020-12-28 | 2021-04-02 | 上海众源网络有限公司 | Object recommendation method, device, equipment and storage medium |
CN112700296A (en) * | 2019-10-23 | 2021-04-23 | 阿里巴巴集团控股有限公司 | Method, device, system and equipment for searching/determining business object |
CN112862539A (en) * | 2021-03-03 | 2021-05-28 | 拉扎斯网络科技(上海)有限公司 | Traffic processing method, traffic processing device, electronic device, storage medium, and program product |
CN113590927A (en) * | 2020-04-30 | 2021-11-02 | 深圳绿米联创科技有限公司 | Content recommendation method, device, server and storage medium |
CN114077701A (en) * | 2020-08-13 | 2022-02-22 | 北京达佳互联信息技术有限公司 | Method and device for determining resource information, computer equipment and storage medium |
Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120029908A1 (en) * | 2010-07-27 | 2012-02-02 | Shingo Takamatsu | Information processing device, related sentence providing method, and program |
CN102542046A (en) * | 2011-12-27 | 2012-07-04 | 纽海信息技术(上海)有限公司 | Book recommendation method based on book contents |
CN104615631A (en) * | 2014-10-29 | 2015-05-13 | 中国建设银行股份有限公司 | Information recommendation method and device |
CN104679835A (en) * | 2015-02-09 | 2015-06-03 | 浙江大学 | Book recommending method based on multi-view hash |
CN105740430A (en) * | 2016-01-29 | 2016-07-06 | 大连理工大学 | Personalized recommendation method with socialization information fused |
CN106649774A (en) * | 2016-12-27 | 2017-05-10 | 北京百度网讯科技有限公司 | Artificial intelligence-based object pushing method and apparatus |
CN106651213A (en) * | 2017-01-03 | 2017-05-10 | 百度在线网络技术(北京)有限公司 | Processing method and device for service orders |
CN106844665A (en) * | 2017-01-20 | 2017-06-13 | 中山大学 | A kind of paper based on the distributed expression of adduction relationship recommends method |
CN106919647A (en) * | 2017-01-17 | 2017-07-04 | 浙江工业大学 | A kind of network structure similitude based on cluster recommends method |
CN107656983A (en) * | 2017-09-08 | 2018-02-02 | 广州索答信息科技有限公司 | A kind of intelligent recommendation method and device based on Application on Voiceprint Recognition |
US20180114235A1 (en) * | 2012-03-13 | 2018-04-26 | American Express Travel Related Services Company, Inc. | Social media distribution of offers based on a consumer relevance value |
-
2018
- 2018-06-12 CN CN201810601391.0A patent/CN110209922B/en active Active
Patent Citations (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120029908A1 (en) * | 2010-07-27 | 2012-02-02 | Shingo Takamatsu | Information processing device, related sentence providing method, and program |
CN102542046A (en) * | 2011-12-27 | 2012-07-04 | 纽海信息技术(上海)有限公司 | Book recommendation method based on book contents |
US20180114235A1 (en) * | 2012-03-13 | 2018-04-26 | American Express Travel Related Services Company, Inc. | Social media distribution of offers based on a consumer relevance value |
CN104615631A (en) * | 2014-10-29 | 2015-05-13 | 中国建设银行股份有限公司 | Information recommendation method and device |
CN104679835A (en) * | 2015-02-09 | 2015-06-03 | 浙江大学 | Book recommending method based on multi-view hash |
CN105740430A (en) * | 2016-01-29 | 2016-07-06 | 大连理工大学 | Personalized recommendation method with socialization information fused |
CN106649774A (en) * | 2016-12-27 | 2017-05-10 | 北京百度网讯科技有限公司 | Artificial intelligence-based object pushing method and apparatus |
CN106651213A (en) * | 2017-01-03 | 2017-05-10 | 百度在线网络技术(北京)有限公司 | Processing method and device for service orders |
CN106919647A (en) * | 2017-01-17 | 2017-07-04 | 浙江工业大学 | A kind of network structure similitude based on cluster recommends method |
CN106844665A (en) * | 2017-01-20 | 2017-06-13 | 中山大学 | A kind of paper based on the distributed expression of adduction relationship recommends method |
CN107656983A (en) * | 2017-09-08 | 2018-02-02 | 广州索答信息科技有限公司 | A kind of intelligent recommendation method and device based on Application on Voiceprint Recognition |
Non-Patent Citations (5)
Title |
---|
YONG ZHENG: "Deviation-based and similarity-based contextual SLIM recommendation algorithms", 《RECSYS \'14: PROCEEDINGS OF THE 8TH ACM CONFERENCE ON RECOMMENDER SYSTEMS》, pages 437 * |
YUE SHI 等: "Mining mood-specific movie similarity with matrix factorization for context-aware recommendation", 《CAMRA \'10: PROCEEDINGS OF THE WORKSHOP ON CONTEXT-AWARE MOVIE RECOMMENDATION》, pages 34 * |
刘冶 等: "基于低秩和局部约束矩阵估计的链接预测方法", 《计算机科学与探索》, vol. 9, no. 03, pages 279 - 291 * |
吴书 等: "情境大数据建模及其在用户行为预测中的应用", 《大数据》, vol. 2, no. 06, pages 110 - 117 * |
王凯: "基于浏览记录的文本挖掘与智能推荐研究与实现", 《中国优秀硕士学位论文全文数据库 信息科技辑》, no. 04, pages 138 - 3865 * |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110737834A (en) * | 2019-10-14 | 2020-01-31 | 腾讯科技(深圳)有限公司 | Business object recommendation method and device, storage medium and computer equipment |
CN110807150A (en) * | 2019-10-14 | 2020-02-18 | 腾讯科技(深圳)有限公司 | Information processing method and device, electronic equipment and computer readable storage medium |
CN110737834B (en) * | 2019-10-14 | 2024-04-02 | 腾讯科技(深圳)有限公司 | Recommendation method and device for business objects, storage medium and computer equipment |
CN112700296B (en) * | 2019-10-23 | 2022-05-27 | 阿里巴巴集团控股有限公司 | Method, device, system and equipment for searching/determining business object |
CN112700296A (en) * | 2019-10-23 | 2021-04-23 | 阿里巴巴集团控股有限公司 | Method, device, system and equipment for searching/determining business object |
CN110955831A (en) * | 2019-11-25 | 2020-04-03 | 北京三快在线科技有限公司 | Article recommendation method and device, computer equipment and storage medium |
CN110955831B (en) * | 2019-11-25 | 2023-04-14 | 北京三快在线科技有限公司 | Article recommendation method and device, computer equipment and storage medium |
CN111079015A (en) * | 2019-12-17 | 2020-04-28 | 腾讯科技(深圳)有限公司 | Recommendation method and device, computer equipment and storage medium |
CN111079015B (en) * | 2019-12-17 | 2021-08-31 | 腾讯科技(深圳)有限公司 | Recommendation method and device, computer equipment and storage medium |
CN111292168B (en) * | 2020-02-06 | 2022-08-02 | 腾讯科技(深圳)有限公司 | Data processing method, device and equipment |
CN111292168A (en) * | 2020-02-06 | 2020-06-16 | 腾讯科技(深圳)有限公司 | Data processing method, device and equipment |
CN111339425A (en) * | 2020-03-05 | 2020-06-26 | 拉扎斯网络科技(上海)有限公司 | Object marking method, device, server and storage medium |
CN111461825A (en) * | 2020-03-30 | 2020-07-28 | 京东数字科技控股有限公司 | Virtual resource generation method and device, electronic equipment and storage medium |
CN111461825B (en) * | 2020-03-30 | 2024-04-09 | 京东科技控股股份有限公司 | Virtual resource generation method and device, electronic equipment and storage medium |
CN113590927A (en) * | 2020-04-30 | 2021-11-02 | 深圳绿米联创科技有限公司 | Content recommendation method, device, server and storage medium |
CN114077701A (en) * | 2020-08-13 | 2022-02-22 | 北京达佳互联信息技术有限公司 | Method and device for determining resource information, computer equipment and storage medium |
CN112215684A (en) * | 2020-10-30 | 2021-01-12 | 腾讯科技(深圳)有限公司 | Clustering method and device for target controllable objects |
CN112215684B (en) * | 2020-10-30 | 2023-07-14 | 腾讯科技(深圳)有限公司 | Clustering method and device for target controllable objects |
CN112507216A (en) * | 2020-12-01 | 2021-03-16 | 北京奇艺世纪科技有限公司 | Data object recommendation method, device, equipment and storage medium |
CN112597395A (en) * | 2020-12-28 | 2021-04-02 | 上海众源网络有限公司 | Object recommendation method, device, equipment and storage medium |
CN112862539A (en) * | 2021-03-03 | 2021-05-28 | 拉扎斯网络科技(上海)有限公司 | Traffic processing method, traffic processing device, electronic device, storage medium, and program product |
Also Published As
Publication number | Publication date |
---|---|
CN110209922B (en) | 2023-11-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110209922A (en) | Object recommendation method, apparatus, storage medium and computer equipment | |
CN113626719B (en) | Information recommendation method, device, equipment, storage medium and computer program product | |
CN110516160A (en) | User modeling method, the sequence of recommendation method of knowledge based map | |
CN111061946A (en) | Scenario content recommendation method and device, electronic equipment and storage medium | |
CN111581510A (en) | Shared content processing method and device, computer equipment and storage medium | |
JP2020523714A (en) | Recommended information acquisition method and device, electronic device | |
CN110162690A (en) | Determine user to the method and apparatus of the interest-degree of article, equipment and storage medium | |
CN113688304A (en) | Training method for search recommendation model, and method and device for sequencing search results | |
US9939272B1 (en) | Method and system for building personalized knowledge base of semantic image segmentation via a selective random field approach | |
CN113256367B (en) | Commodity recommendation method, system, equipment and medium for user behavior history data | |
CN110413888B (en) | Book recommendation method and device | |
CN110222838B (en) | Document sorting method and device, electronic equipment and storage medium | |
CN110580489B (en) | Data object classification system, method and equipment | |
CN114329200B (en) | Personalized self-adaptive binding recommendation model and recommendation method | |
Ben-Shimon et al. | An ensemble method for top-N recommendations from the SVD | |
CN111159563A (en) | Method, device and equipment for determining user interest point information and storage medium | |
CN114896517A (en) | Commodity recommendation method, system, equipment and storage medium | |
CN116186326A (en) | Video recommendation method, model training method, electronic device and storage medium | |
CN117593077A (en) | Prop combination recommendation method and device based on hypergraph neural network and computing equipment | |
CN114817692A (en) | Method, device and equipment for determining recommended object and computer storage medium | |
KR20210016593A (en) | Method and system for recommending product based on artificial intelligence | |
CN117217284A (en) | Data processing method and device | |
CN116975359A (en) | Resource processing method, resource recommending method, device and computer equipment | |
JP5588938B2 (en) | Item recommendation apparatus, method and program | |
CN115600017A (en) | Feature coding model training method and device and media object recommendation method and device |
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 |