CN105469263A

CN105469263A - Commodity recommendation method and device

Info

Publication number: CN105469263A
Application number: CN201410493945.1A
Authority: CN
Inventors: 杨志雄; 梁丽; 张洪
Original assignee: Alibaba Group Holding Ltd
Current assignee: Alibaba Group Holding Ltd
Priority date: 2014-09-24
Filing date: 2014-09-24
Publication date: 2016-04-06

Abstract

The invention discloses a commodity recommendation method which comprises a step of obtaining the historical behavior data of a user browsing commodity, a step of using a pre-generated commodity attribute value prediction model to obtain the currently interested commodity attribute value combination of a user according to the user information of a current user, the historical behavior data and a recommended commodity attribute value combination, and a step of obtaining the commodity set corresponding to the currently interested commodity attribute value combination of the user and recommending at least a part of the commodity in the commodity set to the current user. The invention also discloses a commodity recommendation device. By using the method provided by the invention, the currently interested commodity information can be accurately recommended to the user who browses the commodity at present in real time, the time of browsing the commodity of the user can be saved, the user experience can be effectively improved, and the commodity sale volume of a site can be improved.

Description

A kind of Method of Commodity Recommendation and device

Technical field

The application relates to personalized recommendation field, is specifically related to a kind of Method of Commodity Recommendation.The application provides a kind of device for recommending the commodity simultaneously.

Background technology

Along with the universal of internet and the development of web technology, increasing user selects in online browsing, selects or buy the commodity oneself needed.But along with the quick growth of commodity number and kind, user often requires a great deal of time to carry out browsing and just can find that oneself needs or interested commodity, may cause the continuous loss of the consumer patronizing website like this.In this case, a lot of website adopts various forms of recommended technology to carry out commercial product recommending to user all to some extent, such as, according to Characteristic of Interest and the buying behavior of user, it is recommended to buy or interested commodity and information to user targetedly.Pass through commercial product recommending, the sales volume of shopping website can be improved on the one hand, on the other hand, be convenient to user and find commodity needed for oneself fast, the unnecessary search flow that user simultaneously can also be avoided to produce because frequent search is searched, alleviates the burden of Website server.

In prior art, usually adopt the correlativity calculated between commodity to realize commercial product recommending function, comprise following two steps specifically:

1) calculated off-line often plants the correlativity of commodity.According to user in special time period for the number of times that the various actions (browse, collect, buy) of a certain commodity occur, be weighted summation and obtain the preference numerical value (and can further obtain preference numerical value) of user to these commodity, build the matrix be made up of user, commodity, the preference numerical value (or preference) of user to commodity, then use correlation calculations method calculates the correlation values between two between all commodity, such as, cosine angle formulae as follows is adopted to carry out correlation calculations, wherein R _m,jto represent in matrix m user to the preference numerical value of a jth commodity:

\cos (\overset{&OverBar;}{i}, \overset{&OverBar;}{j}) = \frac{R_{1, i} \times R_{1, j} + R_{2, i} \times R_{1, j} + \cdot \cdot \cdot {+ R}_{m, i} \times R_{m, j}}{\sqrt{{R_{1, i}}^{2} + {R_{2, i}}^{2} + \cdot \cdot \cdot + {R_{m, i}}^{2}} \times \sqrt{{R_{1, j}}^{2} + {R_{2, j}}^{2} + \cdot \cdot \cdot + {R_{m, j}}^{2}}}

According to above-mentioned result of calculation, for each commodity, choose the most relevant to it before K (Topk) commodity for subsequent use as a result.

2) during user's access websites, using the current commodity browsed of this user as benchmark commodity or using just browsed one or more commodity of user as benchmark commodity, according to result of calculation in step 1, choose the multiple commodity relevant to benchmark commodity, show user as recommendation results, further, also multiple benchmark commodity that can be first just browsed to user, (buy according to behavior type, collection or only browse) and time of origin apart from the difference of current time, benchmark commodity are marked, then the multiple commodity relevant to benchmark commodity are chosen according to the result of calculation in step 1, and for each commodity chosen, the score value of the benchmark commodity corresponding with it is multiplied with the correlation values of these commodity with benchmark commodity, product according to obtaining sorts, finally the commodity of the specific quantity come above are showed user as recommendation results.

, there are following two defects in the above-mentioned way of recommendation:

1) demand of user cannot be reflected in real time.The dependent merchandise selected in aforesaid way is that calculated off-line is good, do not adjust for the access behavior that user is nearest, but the buying behavior of user, usually the impact of many real-time key elements can be subject to, such as, user buys electric fan, if Current Temperatures is very high, user may tend to the thermantidote buying band tank; Although if but current be lightning accompanied by peals of thunder in summer, user then may tend to buy band pedestal and heavier stand fan.The purchase intention of user is different, and the commodity browsed on website are recently naturally also different, and the above-mentioned way of recommendation, and owing to not considering user's navigation patterns on line recently, the merchandise news therefore recommending user possibly cannot meet the demand of user.

2) recommendation results is not accurate enough.The above-mentioned way of recommendation is according to the understanding of routine and experience, according to user behavior determination user preference, and the result of calculation of relevance algorithms is directly applied, this mode does not have modeling process, not through training and the checking of great amount of samples data, therefore recommendation results may be caused not accurate enough, and user needs just can find commodity needed for oneself through the browsing of long period, search procedure.

Summary of the invention

The application provides a kind of Method of Commodity Recommendation, with solve prior art cannot for active user in real time, carry out the problem of commercial product recommending exactly.The application provides a kind of device for recommending the commodity in addition.

The application provides a kind of Method of Commodity Recommendation, comprising:

Obtain the historical behavior data that user browses commodity;

According to the user profile of active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user;

Obtain and combine corresponding commodity set with the current interested item property value of described active user, and described active user will be recommended at least partially in the commodity in described commodity set.

Optionally, after the described acquisition of execution combines the step of corresponding commodity set with the current interested item property value of described active user, following operation is performed:

For each commodity in described commodity set, adopt the commodity preference forecast model generated in advance, using the item property value of described user profile, described commodity and described historical behavior data as input, calculate described active user to the interested probability of described commodity;

According to the value order from big to small of described probability, select the commodity of predetermined quantity, and selected commodity are formed new commodity set;

Accordingly, describedly to refer to recommending described active user at least partially in the commodity in described commodity set, recommending described active user at least partially by the commodity in above-mentioned new commodity set.

Optionally, described user profile comprises: user ID, sex, age, income level and address;

Described item property comprises: affiliated classification, price range, style and material.

Optionally, described historical behavior data comprise: presetting the user behavior data in the time period, combine relevant user behavior data to item property value, behavioral data that the user behavior data of being correlated with commodity, user combine item property value and user be to the behavioral data of commodity and relevant scene identity.

Optionally, the described user profile according to active user, described historical behavior data and can for recommend item property value combination, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user, comprising:

Can for the item property value combination recommended for often kind, adopt the item property value prediction model generated in advance, be combined as input with described user profile, described historical behavior data and described item property value, calculate described user and interested probability is combined to described item property value;

Choose the item property value combination that maximal value in above-mentioned probability is corresponding, as the current interested item property value combination of described active user.

Optionally, described acquisition combines corresponding commodity set with the current interested item property value of described active user, realizes in the following way:

The key word that each property value in adopting described item property value to combine is corresponding, retrieves merchandising database, and forms described commodity set with the commodity meeting the combination of described item property value that retrieval obtains.

Optionally, described item property value prediction model generates in the following way:

Adopt LogisticRegression logistic regression algorithm, the historical behavior data of commodity and the combination of item property value is browsed as training set data using user profile, user, set up described item property value prediction model, this model is used for combining interested probability according to the navigation patterns prediction user of user to particular commodity property value;

Described user refers to the combination of particular commodity property value is interested, it is operation by force that user performs the commodity meeting the combination of described item property value, and described is comprise at least one in following element by force: collect, add shopping cart, add receiving tally, purchase.

Optionally, described commodity preference forecast model generates in the following way:

Adopt LogisticRegression logistic regression algorithm, the historical behavior data of commodity and item property value is browsed as training set data using user profile, user, set up described commodity preference forecast model, this model is used for predicting that user is to the interested probability of particular commodity according to the navigation patterns of user;

Described user is interested in particular commodity to be referred to, it is operation by force that user performs described commodity, and described is comprise at least one in following element by force: collect, add shopping cart, add receiving tally, purchase.

Optionally, the process generating described item property value prediction model and the described commodity preference forecast model of generation also comprises the process verified the forecast model set up respectively, and described proof procedure comprises:

Adopt corresponding checking collection data, calculate for assessment of described forecast model whether can desired value;

Judge whether described desired value is greater than the verification threshold preset;

If so, judge that described forecast model may be used for corresponding forecast function; If not, adjust Modling model use data, re-establish corresponding forecast model.

Optionally, described index comprises: AUC index or transaction coverage rate index.。

Accordingly, the application also provides a kind of device for recommending the commodity, comprising:

Historical behavior data capture unit, browses the historical behavior data of commodity for obtaining user;

Item property value combination acquiring unit, for the user profile according to active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user;

Commercial product recommending unit, combines corresponding commodity set for obtaining interested item property value current with described active user, and will recommend described active user at least partially in the commodity in described commodity set;

Described commercial product recommending unit comprises:

Commodity set obtains subelement, combines corresponding commodity set for obtaining interested item property value current with described active user;

Commercial product recommending performs subelement, for recommending described active user at least partially in the commodity in described commodity set.

Optionally, described commercial product recommending unit also comprises:

Commodity probability calculation subelement interested, for for each commodity in described commodity set, adopt the commodity preference forecast model generated in advance, using the item property value of described user profile, described commodity and described historical behavior data as input, calculate described active user to the interested probability of described commodity;

New commodity set obtains subelement, for the order from big to small of the value according to described probability, selects the commodity of predetermined quantity, and selected commodity are formed new commodity set;

Accordingly, described commercial product recommending perform subelement specifically for, recommend described active user at least partially by the commodity in described new commodity set.

Optionally, the user profile that described item property value combination acquiring unit and described commodity probability calculation subelement interested use comprises: user ID, sex, age, income level and address;

The item property that described item property value combination acquiring unit and described commodity probability calculation subelement interested use comprises: affiliated classification, price range, style and material.

Optionally, described item property value combination acquiring unit comprises with the historical behavior data that described commodity probability calculation subelement interested uses: presetting the user behavior data in the time period, combine relevant user behavior data to item property value, and the commodity user behavior data, the user that are correlated with behavioral data that item property value is combined and user to the behavioral data of commodity and relevant scene identity.

Optionally, described item property value combination acquiring unit comprises:

Property value combined probability computation subunit, for combining for the item property value of recommending for often kind, adopt the item property value prediction model generated in advance, be combined as input with described user profile, described historical behavior data and described item property value, calculate described user and interested probability is combined to described item property value;

Property value combination obtains subelement, for choosing item property value combination corresponding to maximal value in above-mentioned probability, as the current interested item property value combination of described active user.

Optionally, described commodity set obtain subelement specifically for, adopt the key word that each property value in the current interested item property value combination of described active user is corresponding, merchandising database is retrieved, and forms described commodity set with the commodity meeting the combination of described item property value that retrieval obtains.

Optionally, described device comprises:

Item property value prediction model generation unit, for adopting LogisticRegression logistic regression algorithm, the historical behavior data of commodity and the combination of item property value is browsed as training set data using user profile, user, set up described item property value prediction model, this model is used for combining interested probability according to the navigation patterns prediction user of user to particular commodity property value.

Optionally, described device comprises:

Commodity preference forecast model generation unit, for adopting LogisticRegression logistic regression algorithm, the historical behavior data of commodity and item property value is browsed as training set data using user profile, user, set up described commodity preference forecast model, this model is used for predicting that user is to the interested probability of particular commodity according to the navigation patterns of user.

Optionally, described item property value prediction model generation unit and described commodity preference forecast model generation unit are except comprising the body subelement for setting up corresponding forecast model respectively, also comprise the corresponding checking subelement for verifying the forecast model set up respectively, corresponding checking subelement comprises:

Index calculate subelement, for adopting corresponding checking collection data, calculate for assessment of described forecast model whether can desired value;

Index judgment sub-unit, for judging whether described desired value is greater than the verification threshold preset; If so, judge that described forecast model may be used for corresponding forecast function; If not, adjust Modling model use data, re-establish corresponding forecast model.

Optionally, the index that described index calculate subelement and described index judgment sub-unit adopt comprises: AUC index or transaction coverage rate index.

Compared with prior art, the application has the following advantages:

The Method of Commodity Recommendation that the application provides, obtain the historical behavior data that user browses commodity, then according to the user profile of active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user, and combine commercial product recommending in corresponding commodity set to described user by with described item property value.Adopt the method that the application provides, the historical behavior data of commodity are browsed according to user, the model set up is utilized to carry out in line computation, thus its current interested merchandise news can be recommended in real time, exactly for the current user browsing commodity, not only can save user browse commodity time, effectively improve Consumer's Experience, the Sales Volume of Commodity of website can also be improved.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the embodiment of a kind of Method of Commodity Recommendation of the application;

Fig. 2 is the schematic diagram of the ROC curve that provides of the embodiment of the present application and AUC index;

Fig. 3 is the processing flow chart obtaining the current interested item property value combination of active user according to item property value prediction model that the embodiment of the present application provides;

Fig. 4 be the embodiment of the present application provide according to commodity preference pattern, from the commodity set obtained, select the processing flow chart of the interested commodity of active user;

Fig. 5 is the schematic diagram of the embodiment of a kind of device for recommending the commodity of the application.

Embodiment

Set forth a lot of detail in the following description so that fully understand the application.But the application can be much different from alternate manner described here to implement, those skilled in the art can when doing similar popularization without prejudice to when the application's intension, and therefore the application is by the restriction of following public concrete enforcement.

In this application, a kind of Method of Commodity Recommendation and a kind of device for recommending the commodity is each provided.Be described in detail one by one in the following embodiments.The technical scheme described due to the present embodiment employs logistic regression algorithm, for the ease of understanding, is first briefly described logistic regression algorithm.

Logistic regression algorithm (LogisticRegression algorithm is called for short LR algorithm), also referred to as Logit model, is one of discrete back-and-forth method model, belongs to multiple variables and analyze category, be generally used for the judgement of classification or the prediction of event occurrence rate.This algorithm is widely used a kind of machine learning algorithm in search field, advertisement promotion field and exemplary application field at present.

Set up the process of Logit model, in fact solve one group of weights (also can be called coefficient) W ₀, W ₁..., W _mprocess, after obtaining above-mentioned weights, when practical application Logit model, according to above-mentioned weights and real data, (that is: dimension is each eigenwert X entering parameter of M ₁-X _m) be weighted summation, obtain corresponding Z value, as follows:

Z=W ₀+ W ₁× X ₁+ ...+W _m× X _m------formula 1

Z in above-mentioned formula 1 also often writes function g (x), then obtains the value of variable P according to the form of sigmoid function as described below:

P (Y = 1 | x) = π (x) = \frac{1}{1 + e^{- g (x)}}

---------formula 2

The codomain of this variable P is [0,1], represent when current enter parameter, there is the probability of (Y=1) in certain event.If for the probable value calculated specifies a discrimination threshold, then this model just can use as sorter.

The key of Modling model as can be seen here, solves each coefficient W in above-mentioned formula 1 exactly ₀, W ₁..., W _mvalue.Below the solution procedure of above-mentioned coefficient is briefly described.

Suppose in training set data, there be n independently training sample { (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), y={0,1}.X wherein _ito be above-mentioned dimension be M enters parameter, each sample (x observed _i, y _i) probability that occurs is:

P(y _i,x _i)＝P(y _i＝1|x _i) ^yi(1-P(y _i＝1|x _i)) ^1-yi

Because each sample is independently, the probability of n sample appearance is exactly the probability multiplication occurred separately, thus the likelihood function of the individual independently sample appearance of n in whole training set data can be obtained, solve each coefficient W when making the likelihood function value in following formula 3 maximum subsequently ₀, W ₁..., W _mvalue.

\max_{w} L (W) = Σ_{i = 1}^{N} [y_{i} \log P (y_{i} = 1 | x_{i}) + (1 - y_{i}) P (y_{i} = 0 | x_{i})]

= Σ_{i = 1}^{N} [y_{i} ({wx}_{i}) - \log (1 + e^{{wx}_{i}})]

----------formula 3

W in above-mentioned formula 3 is exactly the to be solved M dimensional vector comprising W0, W1 ..., WM, xi and yi due to each sample in training set data is known, therefore formula 3 can directly be substituted into, thus obtain one group of non-linear expressions, then the mathematical method such as gradient descent method or newton-La Feisen process of iteration is adopted, solve above-mentioned likelihood function value maximum time each W0, W1 ..., WM parameter value, obtain this group parameter value, the process of establishing of model just completes.Due to process of establishing and the derivation algorithm that relates to of Logit model, be all the prior art of comparative maturity, therefore details do not solved to it and be further described.

Be described above the ultimate principle of logistic regression algorithm, below by embodiment, a kind of Method of Commodity Recommendation provide the application and a kind of device for recommending the commodity are described in detail one by one.

Please refer to Fig. 1, it is the process flow diagram of the embodiment of a kind of Method of Commodity Recommendation of the application.Described method comprises the steps:

Step 101: generate item property value prediction model.

This step adopts LogisticRegression logistic regression algorithm, the historical behavior data of commodity and the combination of item property value is browsed as training set data using user profile, user, set up described item property value prediction model, this model is used for combining interested probability according to the navigation patterns prediction user of user to particular commodity property value.Set up this model and comprise following 5 processes: determine to enter in modeling target, Confirming model each feature, the generating training data collection of parameter, solve W parameter and model is verified, be described in detail respectively below.

1) modeling target is determined.

The common click browse operation of user to commodity can not reflect whether user is interested in the combination of certain item property value well, and user performs by force for operating for the commodity meeting the combination of certain item property value, then usually can illustrate that user is interested in this kind of item property value combination.Described in the application is comprise at least one in following behavior by force: collect, add shopping cart, add receiving tally, place an order buying behavior.

Based on above-mentioned consideration, consider that modeling process needs a large amount of training set data, in an object lesson of the present embodiment, after will being positioned at particular point in time, 7 days users have buying behavior as modeling target under the combination of item property value simultaneously.In other embodiments, also user can be had under the combination of item property value above-mentioned any one by force for as modeling target, equally also can realize the technical scheme of the application.

2) each feature of parameter is entered in Confirming model.

The factor that may affect model prediction result is a lot, and such as: user profile, information attribute value and user browse the historical behavior data etc. of commodity.Wherein, the historical behavior data that the user described in the application browses commodity are not limited to usually said click navigation patterns, are also included in other behaviors occurred in navigation process, such as: collect, add shopping cart, add receiving tally or the buying behavior that places an order.

Specifically choose which data as each feature (that is: the X in formula 1 entering parameter needed for (also need when using a model online use) in modeling process ₁-X _m), depend on the demand of concrete business.

In an object lesson of the present embodiment, choose user profile, item property value combination C _x(representing the combination of various possible item property value) and user browse three class historical behavior data of commodity as each feature entering parameter.Specifically, user profile comprises: user ID, sex, age, income level and address; Item property comprises: affiliated classification, price range, style and material; Three class historical behavior data comprise: user behavior data, combine relevant user behavior data to item property value, behavioral data that user combines item property value, and statistical study is carried out to each class historical behavior data, according to nearest 7 days, 90 days full doses, monthly three dimension statistics, refer to the explanation in table one:

Table one

Each feature in upper table can also further refinement, such as: the number of days etc. of number of visits, the different commodity numbers browsed for nearest 7 days, nearest 7 days customer transaction number of times, nearest 7 days user's collecting commodities number of times, nearest 7 days user's access websites always clicked by the commodity that User behavior in nearest 7 days can also be further subdivided into nearest 7 days; Cx behavior in nearest 7 days can also be further subdivided into the clicked number of times of nearest 7 days these Cx, nearest 7 days these Cx by the number of times that places an order, nearest 7 days these Cx by the transaction number of packages etc. of the dealing money of collection number of times, nearest 7 days these Cx, nearest 7 days these Cx; Cx behavior conversion ratio then refers to that nearest 7 days these Cx are from clicking the conversion ratio (such as: the uv/ that places an order clicks and browses uv) that browsing to places an order buys.

What provide above is the feature that the object lesson of the present embodiment is chosen, other features being different from above-mentioned feature can be chosen in other embodiments, such as: other different from above-mentioned user profile and user-dependent information can be selected as described user profile; The attribute of commodity is also diversified, and other attributes relevant to commodity being different from above-mentioned item property can be selected as described item property; Also the historical behavior data of more horn of plenty can be chosen as the feature entering parameter from the historical behavior data obtained.Due to the one that logistic regression algorithm is in machine learning algorithm, can by the study of a large amount of training set data, training process, identify each feature in parameter of coming in and going out and correlation degree between predicting the outcome, its correlation degree can pass through parameter W _iembody (parameter value of its correspondence of feature that correlation degree is low usually can be smaller, is even 0), done feature as long as therefore can select with the related data of possibility that predict the outcome, participate in modeling process.

3) training set data is generated.

According to 2) in each feature entering parameter of determining, generate multiple training sample, all training samples form training set data jointly.

Under normal circumstances, major part website has and is specifically designed to server or the real-time system that recording user browses the historical behavior data of commodity, the various behavioral datas browsing commodity that the webpage that user shows at client device performs can be uploaded by the mode such as getting ready, usually store with the form of journal file after described server or real-time system receive these data.

The historical behavior data that user browses commodity can obtain usually from above-mentioned journal file, and user profile and information attribute value both can obtain from above-mentioned journal file, also can inquire about acquisition from the database for storing subscriber information and merchandise news.

In the above-mentioned object lesson of the present embodiment, according to 1) in the modeling target and 2 of setting) in each feature entering parameter of selecting, the historical behavior data that user browses commodity in 97 days of specifying are obtained from the journal file of server stores, and described historical behavior data are divided into two parts, data based on the data of first 90 days, the data of latter 7 days are as target data.

According to the historical behavior data of above-mentioned acquisition, generate sample in the following way: according to the feature enumerated in the table one historical behavior extracting data corresponding data of 90 days in the past, using the user profile of extracted data and correspondence, the combination of item property value as the X in sample data _i; For the item property value combination related in this sample data, if there are the data to the commodity generation buying behavior meeting the combination of this item property value in the target data of latter 7 days, illustrate that user is interested in this item property value, the y in so current sample data _i=1, otherwise y _i=0 (referring to formula 3).(each sample data has the X determined to the sample data of employing aforesaid way acquisition sufficient amount _iand y _i), namely form the training set data described in this step.

In the above-mentioned object lesson of the present embodiment, in follow-up 7 days, there is buying behavior as being by force, in other embodiments can also using collection behavior, add the behaviors such as shopping cart as being by force; For choosing of time period, due to 2) in be according to nearest 7 days when selecting feature, 90 days full doses, monthly three dimensions statistics, therefore selected for the 97 heaven-made time periods for generating sample data herein, in other embodiments, the different statistics dimension of selection and different time periods can be needed according to concrete.

4) use training set data, solve W parameter.

By the training set data generated, in the formula 3 that substitution is introduced, and adopt the mathematical method such as gradient descent method or newton-La Feisen process of iteration above, solve likelihood function value of sening as an envoy to maximum time each W ₀, W ₁..., W _mthe value of parameter, namely the item property value prediction model described in this step is obtained, this model can combine interested probability according to the navigation patterns prediction user of user to particular commodity property value, that is: user performs to the commodity meeting the combination of particular commodity property value the probability being by force.

5) model set up is verified.

Theoretically, parameter W has been obtained ₀, W ₁..., W _mdescribed item property value prediction model has just been set up, but as a kind of technical scheme, because in concrete enforcement, the selection strategy of each characteristic dimension is different, the model adopting aforesaid way to set up may be diversified, how judging whether the model set up can meet the forecast demand of technical scheme, with regard to needing, the model set up being verified.The method of checking has multiple, usually adopts calculating AUC to refer to calibration method.

The appearance of AUC (AreaUnderroccurve) index comes from ROC (ReceiverOperatingCharacteristic-recipient's operating characteristic) analytical approach, and the method is the evaluation method about disaggregated model performance.For certain disaggregated model, usually a TPR can be obtained (in all reality as in the sample of positive example according to it in the checking performance integrated on sample, correctly be judged to be the ratio of positive example) and FPR (be in the sample of negative example in all reality, be judged to be the ratio of positive example mistakenly) to put right, this model just can be mapped to a point in ROC plane.The threshold value used when adjusting this category of model, just can obtain a process (0,0), (1,1) curve, that is: the ROC curve of this model, under normal circumstances, this curve all should be in (0,0) and (1,1) top of line, and ROC curve is more drawn close toward upper left side, and the prediction effect of this model is better.

Above-mentioned ROC curve is adopted and is graphically passed judgment on the estimated performance of model, AUC index is then on this basis, the more specifically estimated performance of numerical value to model is adopted to evaluate, specifically, AUC refers to that target value is exactly the size of the part area be in below ROC curve, that is: by (0,0) and (1,1), in the square determined, the size of the area below ROC curve is positioned at.It has been generally acknowledged that, AUC refers to that target value is larger, and the prediction effect of this model is better.Refer to accompanying drawing 2, this accompanying drawing sets forth the ROC curve of two example model, wherein left side ROC area under a curve is 0.84 (AUC=0.84), right side ROC area under a curve is 0.93 (AUC=0.93), the model that right side graph is corresponding as can be seen here, its prediction effect is better than model corresponding to leftmost curve.

According to the definition of above-mentioned AUC index, trapezoidal method is usually adopted to solve the occurrence of this index.Concrete in the present embodiment, can above-mentioned 3) generate training set data while, generate by the checking collection data verifying that sample forms in a large number.In proof procedure, checking collection data substituted in the model established, solve FPR and the TPR numerical value pair under different decision thresholds, each numerical value is to the point of in respective coordinates system, and these points of smooth connection can draw out ROC curve; Then adopt trapezoidal method to solve AUC and refer to target value, specifically, by (FPR, TPR) the adjacent point of numerical value to definition connects with straight line, then through each point, vertical line is done to horizontal ordinate, region segmentation under ROC curve will be become several trapezoidal, solve each trapezoidal area respectively and sue for peace just passable.

Above-described is a kind of computing method of AUC index, in actual applications, some tool software also can be adopted to calculate, such as, AUCCalculator tool software can be used to calculate AUC index.

After solving AUC index, if this refers to that target value is greater than the verification threshold preset, then illustrate that the item property value prediction model set up can apply on line.The described verification threshold preset can be arranged according to the demand of reality, in the above-mentioned object lesson of the present embodiment, the value of this verification threshold is set to 0.7.

If this refers to that target value does not meet above-mentioned requirements, illustrate that predicting the outcome of the item property value prediction model set up can not satisfy the demands, therefore this model can not be applied on line, need to adjust the division entering each feature in parameter and the time dimension of choosing historical behavior data selected in modeling process, and then re-establish this model.

In the above-mentioned object lesson of the present embodiment, AUC index is adopted to verify item property value prediction model, in other embodiments, also transaction coverage rate index (recall) can be adopted to verify model, if the result can not meet the demands, need equally to re-establish this model.

Step 102: generate commodity preference forecast model.

This step adopts LogisticRegression logistic regression algorithm, the historical behavior data of commodity and item property value is browsed as training set data using user profile, user, set up described commodity preference forecast model, this model is used for predicting that user is to the interested probability of particular commodity according to the navigation patterns of user.

Item property value prediction model is similar with generating in step 101, generates commodity preference forecast model and also comprises: determine to enter in modeling target, Confirming model each feature, the generating training data collection of parameter, solve W parameter and carry out such 5 processes of checking to model.Part similar with step 101 in said process, refers to the related description in step 101, below emphasis pair and step 101 difference be described.

1) modeling target is determined.

If it is operation by force that user performs for certain commodity, usually can illustrate that user is interested in these commodity.Described in the application is comprise at least one in following behavior by force: collect, add shopping cart, add receiving tally, place an order buying behavior.

In an object lesson of the present embodiment, after will being positioned at particular point in time, 7 days users have buying behavior as modeling target to commodity.

2) each feature of parameter is entered in Confirming model.

In an object lesson of the present embodiment, choose user profile, item property value, user browse three class historical behavior data of commodity as each feature entering parameter.Specifically, user profile comprises: user ID, sex, age, income level and address; Item property comprises: affiliated classification, price range, style and material; Three class historical behavior data comprise: user behavior data, the user behavior data relevant to commodity, user are to the behavioral data of commodity, and statistical study is carried out to each class historical behavior data, according to nearest 7 days, 90 days full doses, monthly three dimension statistics, refer to the explanation in following table two:

Table two

It should be noted that, in some implementation process, can also using application scenarios mark also as the feature entering parameter, because at different application scenarioss (details page of website homepage, transaction page, the back-stage management page or web site commodity), it may be different that user performs to same displaying merchandise the probability being by force, therefore using this factor also as a feature, the process of establishing of model can be participated in.

3) training set data is generated.

Obtain the historical behavior data that user browses commodity, and according to 2) in each feature entering parameter of determining, from described historical behavior extracting data corresponding data, and jointly generating training sample with corresponding user profile and item property value, all training samples form training set data jointly.

In the above-mentioned object lesson of the present embodiment, similar with the processing procedure in step 101, the historical behavior data of acquisition are divided into two parts, data based on the data of first 90 days, the data of latter 7 days are as target data, when generating sample data: according to the historical behavior extracting data corresponding data of 90 days the pasts of feature of enumerating in table two, by the user profile of extracted data and correspondence and item property value, as the X in sample data _i; For the commodity related in this sample data, if there are the data to this commodity generation buying behavior in the target data of latter 7 days, illustrate that user is interested in these commodity, the y in so current sample data _i=1, otherwise y _i=0 (referring to formula 3).(each sample data has the X determined to the sample data of employing aforesaid way acquisition sufficient amount _iand y _i), namely form the training dataset described in this step.

4) use training set data, solve W parameter.

By the training set data generated, in the formula 3 that substitution is introduced, and adopt the mathematical method such as gradient descent method or newton-La Feisen process of iteration above, solve likelihood function value of sening as an envoy to maximum time each W ₀, W ₁..., W _mthe value of parameter, namely obtains the commodity preference forecast model described in this step, and this model is used for predicting that user is to the interested probability of particular commodity according to the navigation patterns of user, that is: user performs the probability being by force to particular commodity.

5) model set up is verified.

Similar with the respective process in step 101, refer to target value by calculating AUC index or recall, verify the availability of the commodity preference forecast model set up.

Step 103: obtain the historical behavior data that user browses commodity.

After execution of step 101 and 102, establish item property value prediction model and commodity preference forecast model, so far just can apply these two kinds of models on line, for the current user browsing commodity recommends its interested commodity.

Calculating to apply above-mentioned two models on line, needing first to obtain the historical behavior data needed for calculating (obtaining the journal file that such as, can store from the log server of website).In order to the navigation patterns commodity for its recommendation needed for nearest according to active user, the historical behavior data that have recorded the nearest navigation patterns of described active user can be selected, specifically, the historical behavior data described active user can being browsed commodity the same day are included, and the recommendation results calculated like this can reflect the demand that described active user is current relatively in real time.

In the above-mentioned object lesson of the present embodiment, have chosen and comprise the historical behavior data of first 90 days that described active user browses the commodity same day.

Step 104: according to the user profile of active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtains the current interested item property value combination of described active user.

This step adopts the item property value prediction model generated in advance, and obtain the current interested item property value combination of described active user, this processing procedure is divided into following two sub-steps, is described further this two sub-steps below in conjunction with accompanying drawing 3.

Step 104-1: can for the item property value combination recommended for often kind, adopt the item property value prediction model generated in advance, be combined as input with described user profile, described historical behavior data and described item property value, calculate described active user and interested probability is combined to described item property value.

This step will predict that described active user at present may be interested in any item property value combination, therefore the exhaustive current extensive stock property value combination for recommending is needed, and for often kind of item property value combination, adopt described item property value prediction model, calculate described active user to its interested probability.

In the above-mentioned example of the present embodiment, for often kind of item property value combination, by the user profile (this information can obtain from database according to user ID) of described active user, described item property value combination, and the user behavior data corresponding with described active user of the historical behavior extracting data obtained from step 103, the user behavior data that described item property value combination is corresponding, and the user behavior data that described active user combines for described item property value is as each eigenwert entering parameter, substitute in the formula 1 of described item property value prediction model, just can calculate the Z value in formula 1, and then Z value is substituted in corresponding formula 2, just can calculate described active user and at present interested probability be combined to described item property value, that is: described active user performs for the commodity that the combination of this item property value is corresponding the probability that collection or purchase etc. be by force.

Step 104-2: choose the item property value combination that maximal value in above-mentioned probability is corresponding, as the current interested item property value combination of described active user.

In step 104-1, adopt exhaustive mode can combine for the item property value of recommending for often kind, all calculate described active user at present to its interested probability, therefrom choose maximum probable value in this step, using item property value corresponding for this most probable value combination as the current interested item property value combination of described active user.

Step 105: obtain and combine corresponding commodity set with the current interested item property value of described active user.

So far; obtain the current interested item property value combination of described active user; if will the commercial product recommending of this property value combination be met to described active user; user performs by force for the possibility of operation usually can be larger to this kind of commodity; this step is exactly obtain to combine corresponding commodity with described active user interested item property value, for ready to user's Recommendations.

Specifically, querying condition that can be corresponding according to each attribute value generation in described item property value combination, SQL statement is used to inquire about merchandising database, obtain the commodity meeting the combination of described item property value, and all commodity met the demands are formed jointly the commodity set will recommending user.

In concrete enforcement, if website self possesses commercial articles searching engine, then the commercial articles searching engine in station can be directly utilized to obtain the commodity set meeting the combination of item property value.Specifically, by key word corresponding for each property value in described property value combination, submit to the commercial articles searching engine in station, by commercial articles searching engine by the retrieval to inverted index database, obtain corresponding Search Results, that is: meet multiple commodity of described item property value combination, described multiple commodity form the commodity set will recommending user jointly.

Cite a plain example explanation below: item property comprises classification, style, price range, material, carrying out with logistic regression algorithm calculating in the process solved, the value of various different attribute uses digitized representation usually, such as, the current interested item property value of the described active user obtained by step 104 is combined as " 1; 5,3,8 ", wherein classification property value is 1, corresponding " women's dress "; Style attribute is 5, corresponding " Korea Spro's version "; Price range property value is 3, corresponding " price is in 300-500 unit "; Material properties value is 8, corresponding " cotton textiles ", in this case, just following searching request can be submitted to the search engine in station: " classification=women's dress and style=Korea Spro version and price=300-500and material=cotton textiles ", search engine will return the multiple commodity meeting above-mentioned searching request, namely obtains described commodity set.

Certainly, example is above only schematic, and in the specific implementation, item property value may have different values and connotation, and the form passing to the searching request of website search engine may be also different.These can change accordingly according to concrete demand.

After obtaining described commodity set, directly can give described active user by the commercial product recommending in this commodity set, if the commodity amount in described commodity set is many, also can carry out screening (such as selecting front 10 commodity that sales volume is the highest) according to certain strategy, then give described active user by the commercial product recommending after screening.

Step 106: according to the user profile of active user, the item property value of commodity and described historical behavior data, adopt the commodity preference forecast model generated in advance, from described commodity set, choose the interested commodity of described active user, form new commodity set.

Step above have selected according to the interested item property value combination of active user the commodity set recommending user, in order to make recommendation results more accurate, the present embodiment additionally provides a kind of preferred implementation, that is: the commodity preference forecast model generated in advance is adopted, from described commodity set, choose the current interested commodity of described active user, form new commodity set.

Similar with step 104, the processing procedure of this step is also divided into two sub-steps, is described further this two sub-steps below in conjunction with accompanying drawing 4.

Step 106-1: for each commodity in described commodity set, adopt the commodity preference forecast model generated in advance, using the item property value of described user profile, described commodity and described historical behavior data as input, calculate described active user to the interested probability of described commodity.

This step will predict that described active user may be interested in which commodity in described commodity set, therefore needs for each commodity in described commodity set, adopts described commodity preference forecast model, calculate described active user to its interested probability.

In the above-mentioned object lesson of the present embodiment, for each commodity in described commodity set, by the user profile of described active user, the item property value of described commodity, and the user behavior data corresponding with described active user of the historical behavior extracting data obtained from step 103, the user behavior data corresponding with described commodity, and described active user for the user behavior data of described commodity as each eigenwert entering parameter, substitute in the formula 1 of described commodity preference forecast model, just can calculate the Z value in formula 1, and then Z value is substituted in corresponding formula 2, just can calculate described active user to the interested probability of described commodity, that is: described active user performs for these commodity the probability that collection or purchase etc. be by force.

Step 106-2: according to the value order from big to small of described probability, selects the commodity of predetermined quantity, and selected commodity is formed new commodity set.

For often kind of commodity in described commodity set in step 106-1, all calculate described active user to its interested probability, in this step these probable values are arranged according to order from big to small, and according to the quantity preset, therefrom select several commodity be positioned at above, reformulate new commodity set, prepare to recommend described active user.

In an object lesson of the present embodiment, obtain the commodity set that the current interested item property value combination of described active user is corresponding in step 105,100 commodity are comprised in this commodity set, in step 106-1, utilize the commodity preference forecast model generated in advance, calculate this user respectively to these 100 interested probability of commodity, owing to presetting final selection 10 commercial product recommendings to this user, therefore in step 106-2, have chosen 10 commodity that above-mentioned probable value is maximum, constitute new commodity set.

Step 107: give described active user by the commercial product recommending in described new commodity set.

Described active user is given, such as: the commodity details link showing each trade name in this set and correspondence in the page can browsed at present this user by the commercial product recommending in described new commodity set; Or in the page that this user browses at present, show the picture of each commodity in this set and relevant attribute information.Specifically how to recommend to show the commodity in described new commodity set to user, be not the core of technical scheme, technical scheme does not do concrete restriction to this.

Can be found out by description above, by have employed twice logistic regression algorithm in the present embodiment, first time obtains the interested item property value combination of active user according to item property value prediction model, and obtains the commodity set meeting the combination of described item property value; Second time, on primary basis, adopts commodity preferences forecast model, and from described commodity set, selected again active user may several commodity the most interested, and by these commercial product recommendings of finally selecting to user.Owing to have employed logistic regression algorithm Modling model, and be carry out modeling from user's interested item property value combination angle different with the interested commodity of user these two, therefore the recommendation results based on these two models is relatively accurately, user performs to recommended commodity the probability being by force relatively also can be larger, thus effectively can improve the Sales Volume of Commodity of website.

It should be noted that, above-described is the preferred implementation of technical scheme, and wherein some step is not that to realize technical scheme necessary.Such as step 101 and 102, these two steps are modeling process that off-line completes, model is once establish, can Reusability on line within a period of time, in order to reflect the hobby of user more accurately, the historical behavior data regularly can choosing renewal re-establish model, but are not at every turn for current online user carries out the required step of commercial product recommending.In addition, above-described step 106 neither implement the necessary step of technical scheme, this step is to make recommendation results more accurate, so selecting on the basis of Recommendations based on the combination of item property value, again utilize commodity preference pattern to select, in concrete enforcement, do not perform step 106, but the commercial product recommending in the commodity set directly step 105 obtained is to active user, can obtain satisfied recommendation effect equally.If the commodity amount in described commodity set is many, also can carry out screening (such as selecting front 10 commodity that sales volume is the highest) according to certain strategy

In sum, the Method of Commodity Recommendation that the application provides, the historical behavior data of commodity are browsed by obtaining user, then according to the user profile of active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user, and combine commercial product recommending in corresponding commodity set to described user by with described property value.Adopt the method that the application provides, the historical behavior data of commodity are browsed according to user, the model set up is utilized to carry out in line computation, thus its current interested merchandise news can be recommended in real time, exactly for the current user browsing commodity, not only save user browse commodity time, effectively improve Consumer's Experience, the Sales Volume of Commodity of website can also be improved.

In the above-described embodiment, provide a kind of Method of Commodity Recommendation, correspond, the application also provides a kind of device for recommending the commodity.Please refer to Fig. 5, it is the embodiment schematic diagram of a kind of device for recommending the commodity of the application.Because device embodiment is substantially similar to embodiment of the method, so describe fairly simple, relevant part illustrates see the part of embodiment of the method.The device embodiment of following description is only schematic.

A kind of device for recommending the commodity of the present embodiment, comprising: item property value prediction model generation unit 501; Commodity preference forecast model generation unit 502; Historical behavior data capture unit 503; Item property value combination acquiring unit 504; Commodity set acquiring unit 505; Commodity acquiring unit 506 interested; Commercial product recommending performance element 507.

Described item property value prediction model generation unit, for adopting LogisticRegression logistic regression algorithm, the historical behavior data of commodity and the combination of item property value is browsed as training set data using user profile, user, set up described item property value prediction model, this model is used for combining interested probability according to the navigation patterns prediction user of user to particular commodity property value.

Described commodity preference forecast model generation unit, for adopting LogisticRegression logistic regression algorithm, the historical behavior data of commodity and item property value is browsed as training set data using user profile, user, set up described commodity preference forecast model, this model is used for predicting that user is to the interested probability of particular commodity according to the navigation patterns of user.

Described historical behavior data capture unit, browses the historical behavior data of commodity for obtaining user.

Described item property value combination acquiring unit, for the user profile according to active user, described historical behavior data and the item property value combination that can supply recommendation, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user.

Described commodity set acquiring unit, combines corresponding commodity set for obtaining interested item property value current with described active user.

Described commodity acquiring unit interested, for item property value and the described historical behavior data of the user profile according to active user, commodity, adopt the commodity preference forecast model generated in advance, from described commodity set, choose the interested commodity of described active user, form new commodity set.

Described commercial product recommending performance element, for giving described active user by the commercial product recommending in described new commodity set.

Optionally, the user profile that described item property value combination acquiring unit and described commodity acquiring unit interested use comprises: user ID, sex, age, income level and address;

The item property that described item property value combination acquiring unit and described commodity acquiring unit interested use comprises: affiliated classification, price range, style and material.

Optionally, described item property value combination acquiring unit comprises with the historical behavior data that described commodity acquiring unit interested uses: presetting the user behavior data in the time period, combine relevant user behavior data to item property value, and the commodity user behavior data, the user that are correlated with behavioral data that item property value is combined and user to the behavioral data of commodity and relevant scene identity.

Optionally, described item property value combination acquiring unit comprises:

Optionally, described commodity acquiring unit interested comprises:

New commodity set obtains subelement, for the order from big to small of the value according to described probability, selects the commodity of predetermined quantity, and selected commodity are formed new commodity set.

Optionally, described commodity set acquiring unit specifically for, adopt the key word that each property value in the current interested item property value combination of described active user is corresponding, merchandising database is retrieved, and forms described commodity set with the commodity meeting the combination of described item property value that retrieval obtains.

Although the application with preferred embodiment openly as above; but it is not for limiting the application; any those skilled in the art are not departing from the spirit and scope of the application; can make possible variation and amendment, the scope that therefore protection domain of the application should define with the application's claim is as the criterion.

In one typically configuration, computing equipment comprises one or more processor (CPU), input/output interface, network interface and internal memory.

Internal memory may comprise the volatile memory in computer-readable medium, and the forms such as random access memory (RAM) and/or Nonvolatile memory, as ROM (read-only memory) (ROM) or flash memory (flashRAM).Internal memory is the example of computer-readable medium.

1, computer-readable medium comprises permanent and impermanency, removable and non-removable media can be stored to realize information by any method or technology.Information can be computer-readable instruction, data structure, the module of program or other data.The example of the storage medium of computing machine comprises, but be not limited to phase transition internal memory (PRAM), static RAM (SRAM), dynamic RAM (DRAM), the random access memory (RAM) of other types, ROM (read-only memory) (ROM), Electrically Erasable Read Only Memory (EEPROM), fast flash memory bank or other memory techniques, read-only optical disc ROM (read-only memory) (CD-ROM), digital versatile disc (DVD) or other optical memory, magnetic magnetic tape cassette, tape magnetic rigid disk stores or other magnetic storage apparatus or any other non-transmitting medium, can be used for storing the information can accessed by computing equipment.According to defining herein, computer-readable medium does not comprise non-temporary computer readable media (transitorymedia), as data-signal and the carrier wave of modulation.

2, it will be understood by those skilled in the art that the embodiment of the application can be provided as method, system or computer program.Therefore, the application can adopt the form of complete hardware embodiment, completely software implementation or the embodiment in conjunction with software and hardware aspect.And the application can adopt in one or more form wherein including the upper computer program implemented of computer-usable storage medium (including but not limited to magnetic disk memory, CD-ROM, optical memory etc.) of computer usable program code.

Claims

1. a Method of Commodity Recommendation, is characterized in that, comprising:

Obtain the historical behavior data that user browses commodity;

2. Method of Commodity Recommendation according to claim 1, is characterized in that, after the described acquisition of execution combines the step of corresponding commodity set with the current interested item property value of described active user, performs following operation:

3. Method of Commodity Recommendation according to claim 2, is characterized in that, described user profile comprises: user ID, sex, age, income level and address;

4. Method of Commodity Recommendation according to claim 2, it is characterized in that, described historical behavior data comprise: presetting the user behavior data in the time period, combine relevant user behavior data to item property value, behavioral data that the user behavior data of being correlated with commodity, user combine item property value and user be to the behavioral data of commodity and relevant scene identity.

5. Method of Commodity Recommendation according to claim 2, it is characterized in that, the described user profile according to active user, described historical behavior data and can for recommend item property value combination, adopt the item property value prediction model generated in advance, obtain the current interested item property value combination of described active user, comprising:

6., according to the arbitrary described Method of Commodity Recommendation of claim 2-5, it is characterized in that, described acquisition combines corresponding commodity set with the current interested item property value of described active user, realizes in the following way:

7. Method of Commodity Recommendation according to claim 6, is characterized in that, described item property value prediction model generates in the following way:

8. Method of Commodity Recommendation according to claim 7, is characterized in that, described commodity preference forecast model generates in the following way:

9. according to the arbitrary described Method of Commodity Recommendation of claim 7-8, it is characterized in that, the process generating described item property value prediction model and the described commodity preference forecast model of generation also comprises the process verified the forecast model set up respectively, and described proof procedure comprises:

10. Method of Commodity Recommendation according to claim 9, is characterized in that, described index comprises: AUC index or transaction coverage rate index.

11. 1 kinds of devices for recommending the commodity, is characterized in that, comprising:

Described commercial product recommending unit comprises:

12. devices for recommending the commodity according to claim 11, is characterized in that, described commercial product recommending unit also comprises:

13. devices for recommending the commodity according to claim 12, it is characterized in that, the user profile that described item property value combination acquiring unit and described commodity probability calculation subelement interested use comprises: user ID, sex, age, income level and address;

14. devices for recommending the commodity according to claim 12, it is characterized in that, described item property value combination acquiring unit comprises with the historical behavior data that described commodity probability calculation subelement interested uses: presetting the user behavior data in the time period, combine relevant user behavior data to item property value, and the commodity user behavior data, the user that are correlated with behavioral data that item property value is combined and user to the behavioral data of commodity and relevant scene identity.

15. devices for recommending the commodity according to claim 12, is characterized in that, described item property value combination acquiring unit comprises:

16. according to the arbitrary described device for recommending the commodity of claim 12-15, it is characterized in that, described commodity set obtain subelement specifically for, adopt the key word that each property value in the current interested item property value combination of described active user is corresponding, merchandising database is retrieved, and forms described commodity set with the commodity meeting the combination of described item property value that retrieval obtains.

17. devices for recommending the commodity according to claim 16, is characterized in that, comprising:

18. devices for recommending the commodity according to claim 17, is characterized in that, comprising:

19. according to the arbitrary described device for recommending the commodity of claim 17-18, it is characterized in that, described item property value prediction model generation unit and described commodity preference forecast model generation unit are except comprising the body subelement for setting up corresponding forecast model respectively, also comprise the corresponding checking subelement for verifying the forecast model set up respectively, corresponding checking subelement comprises:

20. devices for recommending the commodity according to claim 19, is characterized in that, the index that described index calculate subelement and described index judgment sub-unit adopt comprises: AUC index or transaction coverage rate index.