CN109903138A - A kind of individual commodity recommendation method - Google Patents
A kind of individual commodity recommendation method Download PDFInfo
- Publication number
- CN109903138A CN109903138A CN201910151051.7A CN201910151051A CN109903138A CN 109903138 A CN109903138 A CN 109903138A CN 201910151051 A CN201910151051 A CN 201910151051A CN 109903138 A CN109903138 A CN 109903138A
- Authority
- CN
- China
- Prior art keywords
- multidimensional
- commodity
- user
- information
- property 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 45
- 238000012549 training Methods 0.000 claims description 41
- 230000002452 interceptive effect Effects 0.000 claims description 10
- 230000008569 process Effects 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 239000013065 commercial product Substances 0.000 claims description 4
- 238000004519 manufacturing process Methods 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 230000010365 information processing Effects 0.000 claims 1
- 238000013526 transfer learning Methods 0.000 abstract description 3
- 230000006870 function Effects 0.000 description 16
- 230000009286 beneficial effect Effects 0.000 description 7
- 230000000694 effects Effects 0.000 description 7
- 239000013598 vector Substances 0.000 description 5
- 230000003044 adaptive effect Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002372 labelling Methods 0.000 description 2
- 239000000047 product Substances 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The present invention relates to a kind of individual commodity recommendation methods, comprising: receives the commercial articles searching instruction of user;If user has been classified, the multidimensional for calculating user preference is averaged label information, the recommended models of user be based on multidimensional be averaged label information calculating the first multidimensional property information, the first multidimensional property information is calculated at a distance from the second multidimensional property information of commodity each in merchandise classification, user will be pushed to apart from lesser commodity;Otherwise, the second multidimensional property information based on multiple commodity, calculates the first multidimensional average properties information, and the recommended models based on the first multidimensional average properties information and all types of user determine the recommended models for being suitable for the user.Recommended method provided by the invention first clusters user, and using the Commercial goods labels information and information attribute value of multidimensional, so that recommending accuracy rate high;In addition, by attention transfer learning, the recommended models of new user are formed using existing recommended models, when user is new user to solve the problems, such as cold start-up.
Description
Technical field
The present invention relates to data mining technology fields, more particularly to a kind of individual commodity recommendation method.
Background technique
With the continuous development of network technology, in the epoch of this current information overload, either it is used as information consumer
Or information producer encounters very big challenge: information consumer needs to find oneself from the information of magnanimity interested
Information;Information producer needs to make the information of oneself to show one's talent, and receives the concern of user.Personalized recommendation method can be very well
Solve the above problems.Current main personalized recommendation method mainly includes content-based recommendation method and collaborative filtering
Recommended method, however the generally existing accuracy rate of these methods it is low, cold start-up the problems such as.
Summary of the invention
The present invention provides a kind of individual commodity recommendation method, low to solve recommendation accuracy rate existing in the prior art
The problem of.
The technical scheme to solve the above technical problems is that a kind of individual commodity recommendation method, comprising:
Step 1, the commercial articles searching instruction for receiving demand user, described instruction includes merchandise classification;
Step 2, the storage information based on database, judge whether the demand user has been classified, if so, executing step
3, if it is not, executing step 6;
Step 3, based on the first multidimensional mark for generating the multiple commodity interacted in the demand user and the merchandise classification
Information is signed, the multidimensional for calculating the demand user preference is averaged label information;
Step 4, the recommended models for opening the demand user owning user cluster, the recommended models are based on the multidimensional
Average label information calculates the first multidimensional property information of output;
Step 5, the second multidimensional property for calculating each commodity in the first multidimensional property information and the merchandise classification
The distance between information, and the distance is ranked up according to sequence from small to large, by the preceding preset quantity distance
Corresponding commercial product recommending is to the required user;
Step 6, the second multidimensional property information based on the multiple commodity calculate the first multidimensional average properties letter
Breath, the cluster centre clustered based on the first multidimensional average properties information and each user and recommended models, to the demand
User clusters and determines its recommended models, saves the recommended models, and execute step 3.
The beneficial effects of the present invention are: a kind of individual commodity recommendation algorithm provided by the invention, firstly, being carried out to user
Cluster, each user's cluster is corresponding with one recommended models, so that recommending accuracy rate high;Secondly, utilizing the commodity mark of multidimensional
Information and information attribute value are signed, the feature of commodity can be more comprehensively described, improve the accuracy of recommendation, in addition, when using
When family is new user, the recommended models of new user are formed using recommended models by attention transfer learning, are improved
The convergence rate of the recommended models of new user, and cold start-up can be solved the problems, such as very well, recommendation is high-efficient, and adaptive ability is strong.
Based on the above technical solution, the present invention can also be improved as follows.
Further, the storage information includes multidimensional property information table and multidimensional label information table;
Wherein, the production method of the multidimensional property information table includes:
N number of default commodity in the merchandise classification are preset, N is the positive integer greater than 1;
The the second multidimensional property information and the first multidimensional label information of each default commodity are obtained, and generates institute
State the corresponding multidimensional property information table of N number of default commodity and multidimensional label information table.
It is of the invention further the utility model has the advantages that be previously stored in the database a certain number of default product names and its
Corresponding multidimensional property information and multidimensional label information improve recommendation efficiency convenient for transferring in time in recommendation process.
Further, the generation method of user's cluster includes:
It determines and generated all training users interacted with N number of default commodity, and form each training and use
The commodity list at family, the commodity list include: with the training user generate the n that the interact default commodity and it is wherein each described in
The second multidimensional property information of default commodity, n are the positive integer less than N;
Based on all second multidimensional property information in each commodity list, it is flat to calculate corresponding second multidimensional of the commodity list
Equal attribute information;
The distance between corresponding described second multidimensional average properties information of commodity list described in the every two be less than it is default away from
From when, corresponding two training users of two commodity lists are gathered for one kind, user is formed and clusters and save.
It is of the invention further the utility model has the advantages that the second multidimensional average properties information according to each user clusters, pole
The preference consistency for improving each user's cluster greatly helps to improve recommended models and pushes away to each user in user cluster
Recommend accuracy rate.
Further, each commodity list further include: the first multidimensional label of each default commodity in the commodity list
Information;
Then the training process of the recommended models of each user's cluster includes:
Step 0.1, determine the user cluster in wherein any one of training user;
Step 0.2 determines in the commodity list of the training user wherein any one of default commodity;
This is preset the corresponding first multidimensional label information of commodity and is input to deeply recommendation network by step 0.3, described
Deeply recommendation network exports a third multidimensional property information;
Step 0.4 determines at a distance from the third multidimensional property information from the residue default commodity of the commodity list
The corresponding default commodity of the smallest second multidimensional property information, minimum range are denoted as d,
Step 0.5 passes through state more new formula, is updated, obtains to the first multidimensional label information for presetting commodity
The first new multidimensional label information;
Step 0.6 obtains the score information that the training user presets commodity to this, is based on the d and institute's scoring information,
The loss function value of wheel training is calculated, if the loss function value restrains, completes the training of the recommended models of user cluster,
If it is not, step 0.3 is executed based on the first new multidimensional label information, when pre- in the commodity list to the training user
If commodity have traversed, step 0.1 is executed.
Further beneficial effect of the invention is: deep learning, intensified learning and loss function being combined, effectively increased
The accuracy of the resulting recommended models of training.
Further, the loss function valueIn formula,
It is the third multidimensional property information that next round recycles that target network exports in the deeply recommendation network, wherein being per one-dimensional
Weighted value,It is the third multidimensional property information that the wheel recycles that master network exports in the deeply recommendation network,
It is the third multidimensional property information that next round recycles the master network output, wherein often one-dimensional is weighted value,It is next round
The third multidimensional property information of the master network output is recycled, reward is the reward function value of the wheel, calculation formula
Are as follows:Wherein, rating is institute's scoring information, and e is natural constant;
The state more new formula are as follows: new_s=(1- λ) s+ λ sA, new_s is the first new multidimensional label information,
λ is a constant, and value is [0,1], and s is the first multidimensional label information of the corresponding default commodity of the d, sAFor it is described wherein
First multidimensional label information of one default commodity.
Further beneficial effect of the invention is: for reward function, when distance d is close, scoring is low, then reward is small, pushes away
It is bad to recommend effect;Distance d is remote, scoring is high, then reward is small, and recommendation effect is also bad;And when distance d is close, scoring is high, then recommend
Effect is good.By adding reward function in loss function, the consistent degree of considerations item property that can be fine in training with
Scoring of the user to Recommendations improves the accuracy of training gained recommended models.
Further, described to be gathered based on what the first multidimensional average properties information and each user clustered in the step 6
Class center and recommended models cluster and are determined to the demand user its recommended models, specifically include:
The centre distance for calculating the cluster centre of the first multidimensional average properties information and each user cluster, when wherein
When one centre distance is less than the pre-determined distance, then the demand user is classified as the corresponding user of the centre distance and gathered
Otherwise class is based on each centre distance, calculate the weight that the demand user belongs to each user's cluster, and being based on should
Weight is weighted and averaged the network parameter between the recommended models of all users cluster, obtains pushing away for the demand user
Recommend model.
Further beneficial effect of the invention is: when user is not belonging to any kind, passing through the user and each user
The distance of the central cluster of cluster calculates the weight that the user belongs to each user's cluster, is gathered by weight to by each user
The recommended models of class, which combine, to be weighted and averaged, and the recommended models of the user are obtained, and the cold start-up for avoiding new user is asked
Topic, improves the convergence rate of the recommended models of the new user, effectively improves recommendation efficiency.
Further, before the step 1, the method also includes:
The corresponding all commodity of merchandise classification described in the storage information are clustered;
Then the step 6 specifically includes:
Calculate the second average multidimensional of each commercial articles clustering in the first multidimensional property information and the merchandise classification
The distance between attribute information determines this apart from the smallest commercial articles clustering;
It calculates between the second multidimensional property information of each commodity in the first multidimensional property information and the commercial articles clustering
Distance, and the distance is ranked up according to sequence from small to large, by preceding preset quantity, this is pushed away apart from corresponding commodity
It recommends to the required user.
Further beneficial effect of the invention is: first commodity clustered, the determining commercial articles clustering most preferably recommended, then from
The commodity most preferably recommended are determined in optimal commercial articles clustering, are reduced calculation amount, are improved recommendation efficiency.
Further, after the step 5, the method also includes:
Step 7, the interactive information for obtaining the demand user and recommended commodity generation;
Step 8 is based on the interactive information, determines the first new multidimensional label information of the commodity, and new based on this
First multidimensional label information rebuilds the new recommended models of each user's cluster.
Further beneficial effect of the invention is: according to user to modes such as evaluation, the labellings of recommended commodity
Interactive information regenerates the multidimensional label information of the commodity, prevents the change because of user preference or experience, leads to recommended models
Recommendation inaccuracy, improve the adaptivity of recommended models.
Further, by LDA topic model algorithm, dimension-reduction treatment is carried out to the first multidimensional label information, is obtained new
The first multidimensional label information;
And/or
By LDA topic model algorithm, the first multidimensional property information and the second multidimensional property information are carried out
Dimension-reduction treatment obtains the first new multidimensional property information and the second new multidimensional property information.
Further beneficial effect of the invention is: by dimensionality reduction, computational throughput can be reduced, improve and recommend rate.
The present invention also provides a kind of storage medium, instruction is stored in the storage medium, when computer reads the finger
When enabling, the computer is made to execute individual commodity recommendation method as described in any one of claim 1 to 9.
Detailed description of the invention
Fig. 1 is a kind of flow diagram of individual commodity recommendation method provided by one embodiment of the present invention;
Fig. 2 is the schematic flow chart of the generating mode for user's cluster that another embodiment of the present invention provides;
Fig. 3 is the flow chart element of the training method of the recommended models for each user cluster that another embodiment of the present invention provides
Figure.
Specific embodiment
The principle and features of the present invention will be described below with reference to the accompanying drawings, and the given examples are served only to explain the present invention, and
It is non-to be used to limit the scope of the invention.
Embodiment one
A kind of individual commodity recommendation method 100, as shown in Figure 1, comprising:
Step 110, the commercial articles searching instruction for receiving demand user, instruction includes merchandise classification;
Step 120, the storage information based on database, judge whether demand user has been classified, if so, executing step
130, if it is not, executing step 160;
Step 130, based on the first multidimensional label for generating the multiple commodity interacted in demand user and the merchandise classification
Information, the multidimensional for calculating demand user preference are averaged label information;
Step 140, open demand user owning user cluster recommended models, recommended models be based on multidimensional be averaged label believe
Breath calculates the first multidimensional property information of output;
Step 150, calculate each commodity in the first multidimensional property information and merchandise classification the second multidimensional property information it
Between distance, and distance is ranked up according to sequence from small to large, by preceding preset quantity apart from corresponding commercial product recommending
Give demand user;
Step 160, the second multidimensional property information based on multiple commodity calculate the first multidimensional average properties information, are based on
The cluster centre and recommended models of first multidimensional average properties information and each user cluster, cluster demand user and determine it
Recommended models save the recommended models, and execute step 130.
It should be noted that each multidimensional label information and multidimensional property information are the form of multi-C vector.Multidimensional mark
The specific generating mode for signing information can are as follows: based on interactive information caused by a commodity and its all user, generates the commodity
Multidimensional label information, wherein interactive information can are as follows: user is to information such as the evaluations, scoring, label of commodity.
User's cluster can be carried out by K-Means algorithm, wherein cluster centre is all users couple in each user cluster
The average value for the multidimensional average properties information answered.
In step 130, specifically, each user interacts generation with multiple commodity, multiple commodity are corresponding with multiple multidimensional
Label information obtains the multidimensional of the interested commodity of the user by being weighted and averaged to multiple multidimensional label information
Label information, by the multidimensional expression information input recommended models.
A kind of individual commodity recommendation algorithm provided in this embodiment, firstly, being clustered to user, each user's cluster
It is corresponding with one recommended models, so that recommending accuracy rate high;Secondly, the Commercial goods labels information and item property using multidimensional are believed
Breath, can more comprehensively describe the feature of commodity, improve the accuracy of recommendation, in addition, passing through note when user is new user
Meaning power transfer learning forms the recommended models of new user using recommended models, improves the receipts of the recommended models of new user
Speed is held back, and can solve the problems, such as cold start-up very well, this method recommendation is high-efficient, and adaptive ability is strong.
Preferably, storage information includes: multidimensional property information table and multidimensional label information table;
Wherein, the production method of multidimensional property information table includes:
N number of default commodity in default merchandise classification, N are the positive integer greater than 1;
The the second multidimensional property information and the first multidimensional label information of each default commodity are obtained, and is generated N number of pre-
If the corresponding multidimensional property information table of commodity and multidimensional label information table.
It should be noted that multidimensional label information and multidimensional property information are the form of vector, wherein multidimensional label letter
Each dimensional labels information can be indicated by the commodity with the correlation score of the dimensional labels information in breath, be constituted vector with this
Form, the form of multidimensional property information is the same as multidimensional label information.
In step 130, demand user and the quotient can be determined from the above-mentioned multidimensional label information table stored in database
Category not in generate interaction multiple commodity the first multidimensional label information, be averaged label to calculate the multidimensional of demand user preference
Information;It, can be each in merchandise classification from being determined in the above-mentioned multidimensional property information table stored in database and in step 150
Second multidimensional property information of commodity, carrys out the distance between computation attribute information.
For example, all users have beaten the label information of 1128 dimensions to each film, each film is obtained with 1128
The degree of correlation table of dimension label is as shown in table 1:
Table 1
The degree of correlation of the film native information of each film, as shown in table 2:
It is previously stored with a certain number of default product names and its corresponding multidimensional property information and more in the database
Dimension label information improves recommendation efficiency convenient for transferring in time in recommendation process.
Preferably, the generation method of user's cluster includes:
It determines and generated all training users interacted with N number of default commodity, and form the commodity of each training user
Table, commodity list include: that the n default commodity interacted and wherein more than described the second of each default commodity are generated with the training user
Dimension attribute information, n are the positive integer less than N;
Based on the second multidimensional property information all in each commodity list, calculates corresponding second multidimensional of the commodity list and averagely belong to
Property information;
When the distance between corresponding described second multidimensional average properties information of every two commodity list is less than pre-determined distance,
Corresponding two training users of two commodity lists are gathered for one kind, formation user's cluster.
For example, improve the convergent speed of model in order to improve the accuracy of recommendation, first clustered to user, be
There is the user of identical hobby feature as one kind commodity, targetedly trains, can preferably be applicable in the depth of Mr. Yu class user
Neural network recommendation model.
It is 7 dimension attribute information of film used in user's cluster.Cluster data treatment process can are as follows: finds out first each
Then all films and its 7 dimension attribute information that a user has seen belong to information average value to 7 dimensions of the user, then to all
User is polymerized to K class according to the 7 dimension attribute information acquired, K=12 in this example, and process is as shown in Fig. 2, in figure, A, B, C, D, E, F, G
It indicates the wherein one-dimensional attribute information in 7 dimension attribute information, is finally 12 classes by user's cluster, wherein user 1,7,15 is one
Class, user 2,32 are one kind, and user 4,3,17 is one kind.After user's cluster, so that it may to every a kind of progress recommended models
Training.
It is clustered according to the second multidimensional average properties information of each user, greatly improves the inclined of each user's cluster
Good consistency helps to improve recommended models to the recommendation accuracy rate of each user in user cluster.
Preferably, each commodity list further include: the first multidimensional label information of commodity is each preset in the commodity list;
The then training process of the recommended models of each user's cluster, as shown in Figure 3, comprising:
Step 0.1, determine the user cluster in wherein any one training user;
Step 0.2 determines that wherein any one presets commodity in the commodity list of the training user;
This is preset the corresponding first multidimensional label information of commodity and is input to deeply recommendation network, depth by step 0.3
Strengthen recommendation network and exports a third multidimensional property information;
Step 0.4 presets the determining minimum at a distance from the third multidimensional property information in commodity from the residue of the commodity list
The corresponding default commodity of the second multidimensional property information, minimum range is denoted as d;
Step 0.5 passes through state more new formula, is updated, obtains to the first multidimensional label information for presetting commodity
The first new multidimensional label information;
Step 0.6 obtains the score information that the training user presets commodity to this, is based on d and institute's scoring information, calculates
The loss function value of wheel training completes the training of the recommended models of user cluster, if it is not, base if loss function value restrains
In the first new multidimensional label information, step 0.3 is executed, has traversed, has held when presetting commodity in the commodity list to the training user
Row step 0.1.
Theoretically intensified learning is being trained always until model is restrained.When the user of all categories have one to one's name
Recommended models after, the training stage terminates.Deep learning, intensified learning and loss function are combined, trained institute is effectively increased
The accuracy of the recommended models obtained.
Preferably, the loss function valueIn formula,
It is the third multidimensional property information that next round recycles that target network exports in the deeply recommendation network, wherein being per one-dimensional
Weighted value,It is the third multidimensional property information that the wheel recycles that master network exports in the deeply recommendation network,
It is the third multidimensional property information that next round recycles the master network output, wherein often one-dimensional is weighted value,It is next round
The third multidimensional property information of the master network output is recycled, reward is the reward function value of the wheel, calculation formula
Are as follows:Wherein, rating is institute's scoring information, and e is natural constant;
The state more new formula are as follows: new_s=(1- λ) s+ λ sA, new_s is the first new multidimensional label information,
λ is a constant, and value is [0,1], and s is the first multidimensional label information of the corresponding default commodity of the d, sAFor it is described wherein
First multidimensional label information of one default commodity.
For reward function, when distance d is close, scoring is low, then reward is small, and recommendation effect is bad;Distance d is remote, scoring is high,
Then reward is small, and recommendation effect is also bad;And when distance d is close, scoring is high, then recommendation effect is good.By adding in loss function
Add reward function, scoring of the consistent degree and user of consideration item property that can be fine in training to Recommendations improves
The accuracy of training gained recommended models.
Preferably, it in step 160, the cluster centre that is clustered based on the first multidimensional average properties information and each user and pushes away
Model is recommended, its recommended models is clustered and determined to demand user, is specifically included:
The centre distance for calculating the cluster centre of the first multidimensional average properties information and each user cluster, when one of them
When the centre distance is less than the pre-determined distance, then the demand user is classified as the corresponding user of the centre distance and clustered,
Otherwise, it is based on each centre distance, the demand user is calculated and belongs to the weight of each user's cluster, and be based on the power
Weight is weighted and averaged the network parameter between the recommended models of all users cluster, obtains the recommendation of the demand user
Model.
Cluster centre is the average value of the corresponding multidimensional average properties information of all users in each user cluster.
When user is not belonging to any kind, through the user at a distance from the central cluster that each user clusters, calculate
The user belongs to the weight of each user's cluster, is combined by weight the recommended models for clustering each user and is added
Weight average obtains the recommended models of the user, avoids the cold start-up problem of new user, improves the recommended models of the new user
Convergence rate, effectively improve recommendation efficiency.
Assuming that a shared n user cluster, wherein diIt is the cluster centre that demand user distance has i-th of user cluster
The distance between.
So, the network parameter W of demand user recommended modelsnew=wTW, w are one-dimensional weight vectors, wiFor each of w
Column, represent the weight of i-th of recommended models, and W is n dimension, per the one-dimensional network parameter for clustering corresponding recommended models by n user
The network parameter vector constituted.
Preferably, before the step 1, method 100 further include:
The corresponding all commodity of merchandise classification in storage information are clustered;
Then step 160 specifically includes:
Calculate the second average multidimensional of each commercial articles clustering in the first multidimensional property information and the merchandise classification
The distance between attribute information determines this apart from the smallest commercial articles clustering;
It calculates between the second multidimensional property information of each commodity in the first multidimensional property information and the commercial articles clustering
Distance, and the distance is ranked up according to sequence from small to large, by preceding preset quantity, this is pushed away apart from corresponding commodity
It recommends to the required user.
First commodity are clustered, determine the commercial articles clustering most preferably recommended, then are determined most preferably from optimal commercial articles clustering
The commodity of recommendation, reduce calculation amount, improve recommendation efficiency.
Preferably, after step 150, method 100 further include:
The interactive information that acquisition demand user and recommended commodity generate;
Based on interactive information, the first new multidimensional label information of the commodity is determined, and based on the first new multidimensional mark
Information is signed, the new recommended models of each user's cluster are rebuild.
Like changing according to user, the cluster and each user for adjusting user cluster corresponding recommended models, using most
Accurate recommended models are done for user and are recommended, improve adaptive recommendation ability and individual commodity recommendation effect.
According to user to the interactive information of the modes such as evaluation, the labelling of recommended commodity, the commodity are regenerated
Multidimensional label information prevents the change because of user preference or experience, leads to the recommendation inaccuracy of recommended models, improves recommendation
The adaptivity of model.
It should be noted that can also be clustered again to user in real time according to the hobby of user.
Preferably, by LDA topic model algorithm, dimension-reduction treatment is carried out to the first multidimensional label information, is obtained new
The first multidimensional label information;
And/or
By LDA topic model algorithm, the first multidimensional property information and the second multidimensional property information are carried out
Dimension-reduction treatment obtains the first new multidimensional property information and the second new multidimensional property information.
LDA (Latent Dirichlet Allocation) is that a kind of document subject matter generates model, also referred to as one three layers
Bayesian probability model.By dimensionality reduction, computational throughput can be reduced, improve and recommend rate.
Embodiment two
A kind of storage medium is stored with instruction in storage medium, when computer reads described instruction, executes computer
Such as the individual commodity recommendation method of embodiment one.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (10)
1. a kind of individual commodity recommendation method characterized by comprising
Step 1, the commercial articles searching instruction for receiving demand user, described instruction includes merchandise classification;
Step 2, the storage information based on database, judge whether the demand user has been classified, if so, step 3 is executed, if
It is no, execute step 6;
Step 3 is believed based on the first multidimensional label for generating the multiple commodity interacted in the demand user and the merchandise classification
Breath, the multidimensional for calculating the demand user preference are averaged label information;
Step 4, the recommended models for opening the demand user owning user cluster, the recommended models are average based on the multidimensional
Label information calculates the first multidimensional property information of output;
Step 5, the second multidimensional property information for calculating each commodity in the first multidimensional property information and the merchandise classification
The distance between, and the distance is ranked up according to sequence from small to large, the preceding preset quantity distance is corresponding
Commercial product recommending give the demand user;
Step 6, the second multidimensional property information based on the multiple commodity calculate the first multidimensional average properties information, base
It is poly- to the demand user in the cluster centre and recommended models of the first multidimensional average properties information and each user cluster
Class simultaneously determines its recommended models, saves the recommended models and executes step 3.
2. a kind of individual commodity recommendation method according to claim 1, which is characterized in that the storage information includes more
Dimension attribute information table and multidimensional label information table;
Wherein, the production method of the multidimensional property information table includes:
N number of default commodity in the merchandise classification are preset, N is the positive integer greater than 1;
The the second multidimensional property information and the first multidimensional label information of each default commodity are obtained, and generates the N
The corresponding multidimensional property information table of a default commodity and multidimensional label information table.
3. a kind of individual commodity recommendation method according to claim 2, which is characterized in that the generation of user's cluster
Method includes:
It determines and generated all training users interacted with N number of default commodity, and form each training user's
Commodity list, the commodity list include: to generate the n that the interact default commodity with the training user and wherein each described preset
The second multidimensional property information of commodity, n are the positive integer less than N;
Based on all second multidimensional property information in each commodity list, calculates corresponding second multidimensional of the commodity list and averagely belong to
Property information;
When the distance between corresponding described second multidimensional average properties information of commodity list described in the every two is less than pre-determined distance,
Corresponding two training users of two commodity lists are gathered for one kind, user is formed and clusters and save.
4. a kind of individual commodity recommendation method according to claim 3, which is characterized in that each commodity list also wraps
It includes: the first multidimensional label information of each default commodity in the commodity list;
Then the training process of the recommended models of each user's cluster includes:
Step 0.1, determine the user cluster in wherein any one of training user;
Step 0.2 determines in the commodity list of the training user wherein any one of default commodity;
This is preset the corresponding first multidimensional label information of commodity and is input to deeply recommendation network, the depth by step 0.3
Strengthen recommendation network and exports a third multidimensional property information;
Step 0.4 determines the minimum at a distance from the third multidimensional property information from the residue default commodity of the commodity list
The corresponding default commodity of the second multidimensional property information, minimum range is denoted as d;
Step 0.5 passes through state more new formula, is updated, obtains new to the first multidimensional label information for presetting commodity
First multidimensional label information;
Step 0.6 obtains the score information that the training user presets commodity to this, is based on the d and institute's scoring information, calculates
The loss function value of wheel training completes the training of the recommended models of user cluster if the loss function value restrains, if
It is no, based on the first new multidimensional label information, step 0.3 is executed, when default in the commodity list to the training user
Commodity have traversed, and execute step 0.1.
5. a kind of individual commodity recommendation method according to claim 4, which is characterized in that the loss function valueIn formula,It is that next round recycles the deeply recommendation
The third multidimensional property information that target network exports in network, wherein often one-dimensional is weighted value,It is that the wheel recycles the depth
Degree strengthens the third multidimensional property information of master network output in recommendation network,It is that next round recycles the master network and exports
Third multidimensional property information, wherein often one-dimensional is weighted value,It is the third multidimensional that next round recycles the master network output
Attribute information, reward are the reward function value of the wheel, its calculation formula is:Wherein, rating is institute's scoring information, and e is natural constant;
The state more new formula are as follows: new_s=(1- λ) s+ λ sA, new_s is the first new multidimensional label information, λ mono-
A constant, value are [0,1], and s is the first multidimensional label information of the corresponding default commodity of the d, sAFor it is described one of them
First multidimensional label information of the default commodity.
6. a kind of individual commodity recommendation method according to claim 3, which is characterized in that in the step 6, the base
It is poly- to the demand user in the cluster centre and recommended models of the first multidimensional average properties information and each user cluster
Class simultaneously determines its recommended models, specifically includes:
The centre distance for calculating the cluster centre of the first multidimensional average properties information and each user cluster, when one of them
When the centre distance is less than the pre-determined distance, then the demand user is classified as the corresponding user of the centre distance and clustered,
Otherwise, it is based on each centre distance, the demand user is calculated and belongs to the weight of each user's cluster, and be based on the power
Weight is weighted and averaged the network parameter between the recommended models of all users cluster, obtains the recommendation of the demand user
Model.
7. a kind of individual commodity recommendation method according to any one of claims 1 to 6, which is characterized in that the step 1
Before, the method also includes:
The corresponding all commodity of merchandise classification described in the storage information are clustered;
Then the step 6 specifically includes:
Calculate the second average multidimensional property of each commercial articles clustering in the first multidimensional property information and the merchandise classification
The distance between information determines this apart from the smallest commercial articles clustering;
Calculate between the second multidimensional property information of each commodity in the first multidimensional property information and the commercial articles clustering away from
From and the distance being ranked up according to sequence from small to large, by preceding preset quantity, this is given apart from corresponding commercial product recommending
User needed for described.
8. a kind of individual commodity recommendation method according to any one of claims 1 to 6, which is characterized in that the step 5
Later, the method also includes:
Obtain the interactive information of the demand user and recommended commodity generation;
Based on the interactive information, the first new multidimensional label information of the commodity is determined, and based on the first new multidimensional mark
Information is signed, the new recommended models of each user's cluster are rebuild.
9. a kind of individual commodity recommendation method according to any one of claims 1 to 6, which is characterized in that pass through LDA master
Model algorithm is inscribed, dimension-reduction treatment is carried out to the first multidimensional label information, obtains the first new multidimensional label information;
And/or
By LDA topic model algorithm, dimensionality reduction is carried out to the first multidimensional property information and the second multidimensional property information
Processing obtains the first new multidimensional property information and the second new multidimensional property information.
10. a kind of storage medium, which is characterized in that instruction is stored in the storage medium, when computer reads described instruction
When, so that the computer is executed individual commodity recommendation method as described in any one of claim 1 to 9.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910151051.7A CN109903138B (en) | 2019-02-28 | 2019-02-28 | Personalized commodity recommendation method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910151051.7A CN109903138B (en) | 2019-02-28 | 2019-02-28 | Personalized commodity recommendation method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109903138A true CN109903138A (en) | 2019-06-18 |
CN109903138B CN109903138B (en) | 2021-05-18 |
Family
ID=66945757
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910151051.7A Expired - Fee Related CN109903138B (en) | 2019-02-28 | 2019-02-28 | Personalized commodity recommendation method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109903138B (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111047394A (en) * | 2019-11-20 | 2020-04-21 | 泰康保险集团股份有限公司 | Commodity recommendation method and device, electronic equipment and computer readable medium |
CN111079001A (en) * | 2019-11-26 | 2020-04-28 | 贝壳技术有限公司 | Decoration recommendation information generation method and device, storage medium and electronic equipment |
CN111488524A (en) * | 2020-04-08 | 2020-08-04 | 吉林大学 | Attention-oriented semantic-sensitive label recommendation method |
CN112597392A (en) * | 2020-12-25 | 2021-04-02 | 厦门大学 | Recommendation system based on dynamic attention and hierarchical reinforcement learning |
CN113988207A (en) * | 2021-11-09 | 2022-01-28 | 长春理工大学 | Article recommendation method and system |
CN116797322A (en) * | 2023-08-16 | 2023-09-22 | 杭州阿里巴巴海外网络科技有限公司 | Method for providing commodity object information and electronic equipment |
Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103077220A (en) * | 2012-12-29 | 2013-05-01 | 中国科学院深圳先进技术研究院 | User group correlation degree-based personalized recommendation method and system |
US20150269338A1 (en) * | 2009-12-09 | 2015-09-24 | Jonathan Kaleb Adams | Computer System for Medical Triage Determinations and Related System Interactions |
CN105608219A (en) * | 2016-01-07 | 2016-05-25 | 上海通创信息技术有限公司 | Stream-oriented recommended engine, recommendation system and recommendation method based on clustering |
CN106779946A (en) * | 2016-12-16 | 2017-05-31 | Tcl集团股份有限公司 | A kind of film recommends method and device |
CN106940801A (en) * | 2016-01-04 | 2017-07-11 | 中国科学院声学研究所 | A kind of deeply for Wide Area Network learns commending system and method |
CN108205682A (en) * | 2016-12-19 | 2018-06-26 | 同济大学 | It is a kind of for the fusion content of personalized recommendation and the collaborative filtering method of behavior |
CN108429776A (en) * | 2017-02-13 | 2018-08-21 | 阿里巴巴集团控股有限公司 | Method for pushing, device, client, interactive device and the system of network object |
CN108960945A (en) * | 2017-05-18 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Method of Commodity Recommendation and device |
CN109033453A (en) * | 2018-08-24 | 2018-12-18 | 安徽大学 | RBM and differential privacy protection based clustering movie recommendation method and system |
CN109166017A (en) * | 2018-10-12 | 2019-01-08 | 平安科技(深圳)有限公司 | Method for pushing, device, computer equipment and storage medium based on reunion class |
-
2019
- 2019-02-28 CN CN201910151051.7A patent/CN109903138B/en not_active Expired - Fee Related
Patent Citations (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150269338A1 (en) * | 2009-12-09 | 2015-09-24 | Jonathan Kaleb Adams | Computer System for Medical Triage Determinations and Related System Interactions |
CN103077220A (en) * | 2012-12-29 | 2013-05-01 | 中国科学院深圳先进技术研究院 | User group correlation degree-based personalized recommendation method and system |
CN106940801A (en) * | 2016-01-04 | 2017-07-11 | 中国科学院声学研究所 | A kind of deeply for Wide Area Network learns commending system and method |
CN105608219A (en) * | 2016-01-07 | 2016-05-25 | 上海通创信息技术有限公司 | Stream-oriented recommended engine, recommendation system and recommendation method based on clustering |
CN106779946A (en) * | 2016-12-16 | 2017-05-31 | Tcl集团股份有限公司 | A kind of film recommends method and device |
CN108205682A (en) * | 2016-12-19 | 2018-06-26 | 同济大学 | It is a kind of for the fusion content of personalized recommendation and the collaborative filtering method of behavior |
CN108429776A (en) * | 2017-02-13 | 2018-08-21 | 阿里巴巴集团控股有限公司 | Method for pushing, device, client, interactive device and the system of network object |
CN108960945A (en) * | 2017-05-18 | 2018-12-07 | 北京京东尚科信息技术有限公司 | Method of Commodity Recommendation and device |
CN109033453A (en) * | 2018-08-24 | 2018-12-18 | 安徽大学 | RBM and differential privacy protection based clustering movie recommendation method and system |
CN109166017A (en) * | 2018-10-12 | 2019-01-08 | 平安科技(深圳)有限公司 | Method for pushing, device, computer equipment and storage medium based on reunion class |
Non-Patent Citations (7)
Title |
---|
KEUNCHAN PARK 等: "Deep Neural Networks for News Recommendations", 《PROCEEDINGS OF THE 2017 ACM ON CONFERENCE ON INFORMATION AND KNOWLEDGE MANAGEMENT》 * |
WEI WEI 等: "Learning to Find Topic Experts in Twitter via Different Relations", 《IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING 》 * |
XIAOHUI LI 等: "Using Multidimensional Clustering Based Collaborative Filtering Approach Improving Recommendation Diversity", 《2012 IEEE/WIC/ACM INTERNATIONAL CONFERENCES ON WEB INTELLIGENCE AND INTELLIGENT AGENT TECHNOLOGY》 * |
梁丽君: "基于用户属性聚类的协同过滤推荐算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
王科强: "基于矩阵分解的个性化推荐***", 《中国博士学位论文全文数据库 信息科技辑》 * |
王雯: "基于深度强化学习的引导式习题推荐模型研究", 《中国优秀硕士学位论文全文数据库 经济与管理科学辑》 * |
闫昶: "基于协同过滤的景点推荐WebGIS平台设计与实现", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111047394A (en) * | 2019-11-20 | 2020-04-21 | 泰康保险集团股份有限公司 | Commodity recommendation method and device, electronic equipment and computer readable medium |
CN111079001A (en) * | 2019-11-26 | 2020-04-28 | 贝壳技术有限公司 | Decoration recommendation information generation method and device, storage medium and electronic equipment |
CN111488524A (en) * | 2020-04-08 | 2020-08-04 | 吉林大学 | Attention-oriented semantic-sensitive label recommendation method |
CN111488524B (en) * | 2020-04-08 | 2022-08-16 | 吉林大学 | Attention-oriented semantic-sensitive label recommendation method |
CN112597392A (en) * | 2020-12-25 | 2021-04-02 | 厦门大学 | Recommendation system based on dynamic attention and hierarchical reinforcement learning |
CN113988207A (en) * | 2021-11-09 | 2022-01-28 | 长春理工大学 | Article recommendation method and system |
CN116797322A (en) * | 2023-08-16 | 2023-09-22 | 杭州阿里巴巴海外网络科技有限公司 | Method for providing commodity object information and electronic equipment |
CN116797322B (en) * | 2023-08-16 | 2024-02-09 | 杭州阿里巴巴海外网络科技有限公司 | Method for providing commodity object information and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN109903138B (en) | 2021-05-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109903138A (en) | A kind of individual commodity recommendation method | |
CN103377296B (en) | A kind of data digging method of many indexs evaluation information | |
CN108920527A (en) | A kind of personalized recommendation method of knowledge based map | |
CN108334592A (en) | A kind of personalized recommendation method being combined with collaborative filtering based on content | |
CN109064285A (en) | A kind of acquisition commercial product recommending sequence and Method of Commodity Recommendation | |
CN108665323A (en) | A kind of integrated approach for finance product commending system | |
Wang et al. | Social recommendation with optimal limited attention | |
CN109711925A (en) | Cross-domain recommending data processing method, cross-domain recommender system with multiple auxiliary domains | |
CN103605723B (en) | A kind of video recommendation method based on particle cluster algorithm | |
CN106897914A (en) | A kind of Method of Commodity Recommendation and system based on topic model | |
CN114399251B (en) | Cold-chain logistics recommendation method and device based on semantic web and cluster preference | |
CN107545471A (en) | A kind of big data intelligent recommendation method based on Gaussian Mixture | |
CN111310038B (en) | Information recommendation method and device, electronic equipment and computer-readable storage medium | |
CN104657336B (en) | A kind of personalized recommendation method based on half cosine function | |
Hossein Javaheri | Response modeling in direct marketing: a data mining based approach for target selection | |
CN106776859A (en) | Mobile solution App commending systems based on user preference | |
CN105844302A (en) | Depth-learning-based method for automatically calculating commodity trend indexes | |
CN110334278A (en) | A kind of web services recommended method based on improvement deep learning | |
CN115374288A (en) | Recommendation method based on knowledge graph and attention mechanism | |
CN111815403A (en) | Commodity recommendation method and device and terminal equipment | |
CN114723535A (en) | Supply chain and knowledge graph-based item recommendation method, equipment and medium | |
Puraram et al. | Thai food recommendation system using hybrid of particle swarm optimization and K-means algorithm | |
CN108491477A (en) | Neural network recommendation method based on multidimensional cloud and user's dynamic interest | |
Raeder et al. | Scalable supervised dimensionality reduction using clustering | |
CN111612583A (en) | Individualized shopping guide system based on clustering |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20210518 |
|
CF01 | Termination of patent right due to non-payment of annual fee |