CN106407477A - Multidimensional interconnection recommendation method and system - Google Patents

Abstract

The invention discloses a multidimensional interconnection recommendation method and system. An offline recommendation result is generated according to the historical behavior log of a user, then a near-line recommendation result is generated according to the real-time behavior log of the user, and finally the two recommendation results are fused to generate a final recommendation result, so that the recommendation precision is high by using the advantages of different sorting algorithms, the individuation requirement of the user can be met, the content to which the user pay close attention is controlled more accurately, and the system has high adaptability.

Description

Multidimensional interconnection recommendation method and system

Technical Field

The invention belongs to the technical field of networks, and particularly relates to a multidimensional interconnection recommendation method and system.

Background

At present, a mass of information is screened and filtered by a media website, and information which is most concerned by a user and is most interesting is displayed in front of the user by utilizing a search engine. The current popular recommendation algorithms are single and can only meet the requirement of local recommendation.

At present, a media website of an e-commerce platform screens and filters massive information, when a search engine is used for screening the information, when a user cannot accurately describe the demand effect of the user, the demand effect is greatly discounted, the versatility of the user cannot be met, personalized demands are difficult to fully embody, the limitations of various recommendation methods are exposed, the algorithm A for a certain scene has a good effect, and the algorithm B for another scene has a good effect.

Disclosure of Invention

The invention aims to: in order to solve the problems in the prior art and provide more accurate recommendation information for users, the invention provides a multi-dimensional interconnection recommendation method and system.

The technical scheme of the invention is as follows: a multi-dimensional interconnection recommendation method comprises the following steps:

s1, extracting a user historical behavior log to form basic data, and generating an offline recommendation result by adopting an Ensemble training method;

s2, receiving the real-time behavior log of the user, and generating a near-line recommendation result by adopting a linear fusion method according to the real-time action of the user;

and S3, fusing the offline recommendation result in the step S1 with the online recommendation result in the step S2, and generating a recommendation result by adopting a priority algorithm.

Further, the Ensemble training method comprises an L1 layer classifier and an L2 layer classifier.

Further, the step S1 is to extract a user history behavior log to form basic data, and the step of generating an offline recommendation result by using an Ensemble training method specifically includes the following sub-steps:

s11, taking the basic data as a training sample and dividing the training sample into a Trainpig part and a Test Pig part according to a ratio of 1: 1;

s12, extracting features of the Train pig part in the step S11 to form feature vectors, and training an L1-layer classification model;

s13, predicting the Test Pig part in the step S11 by adopting the L1 layer classification model in the step S12, and generating a L1 layer prediction result;

s14, generating an input feature vector of an L2 layer according to the L1 layer prediction result in the step S13, and training an L2 layer classification model for the Test Pig part;

s15, retraining the L1-layer classification model for all training samples;

and S16, extracting features of the sample to be tested, generating two prediction results by respectively adopting the L1-layer classification model in the step S12 and the L1-layer classification model retrained in the step S15, and generating an offline recommendation result according to the two prediction results by adopting the L2-layer classification model in the step S14.

Further, the behavior log includes a presentation log and a click log.

Further, the presentation log comprises presented items, algorithms adopted when the items are recommended, recommended positions and item corresponding weights.

Further, the step S2 of generating the nearline recommendation result by using a linear fusion method according to the real-time action of the user specifically includes: and updating the weight of the recommended strategy corresponding to the item according to the real-time action behavior of the user, and recalculating the recommendation result.

Further, if the real-time action behavior of the user is a presentation log, the weight of the recommended strategy corresponding to the item is reduced, and the calculation formula is represented as:

wherein, β'_kFor updated weights, β_kAs the original weight, Ctr_iTo show the average click rate for position i, Score_kScore for algorithm K to this item_itemAnd lambda is the decay constant of the position click rate and the decay constant of the algorithm click rate, wherein lambda is the total score of the item.

Further, if the real-time action behavior of the user is a click log, the weight of the recommended strategy corresponding to the item is increased, and the calculation formula is represented as:

wherein,is the click decay constant.

In order to further explain the multidimensional interconnection recommendation method of the present invention, the present invention further provides a multidimensional interconnection recommendation system, including:

the off-line subsystem is used for establishing a training sample according to the historical behavior log of the user, generating an off-line recommendation result and sending the off-line recommendation result to the on-line subsystem;

the online system comprises an online subsystem and an online subsystem, wherein the online subsystem is used for receiving a real-time behavior log of a user, generating an online recommendation result according to the real-time behavior log of the user and sending the online recommendation result to the online subsystem;

and the online subsystem is used for receiving and fusing the offline recommendation result generated by the offline subsystem and the online recommendation result generated by the online subsystem, and feeding the fused recommendation result back to the user.

Furthermore, the near line subsystem is arranged at a service end, and the online subsystem faces to a user end.

The invention has the beneficial effects that: according to the method, the offline recommendation result is generated according to the historical behavior log of the user, the online recommendation result is generated according to the real-time behavior log of the user, the two recommendation results are fused to generate the final recommendation result, the advantages of different classification algorithms are utilized, the recommendation precision is high, the personalized requirements of the user can be met, the content concerned by the user can be mastered more accurately, and the method has good adaptability.

Drawings

FIG. 1 is a flow chart of a multidimensional interconnection recommendation method of the present invention.

FIG. 2 is a schematic structural diagram of the multi-dimensional interconnection recommendation system of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Fig. 1 is a schematic flow chart of a multidimensional interconnection recommendation method according to the present invention. A multi-dimensional interconnection recommendation method comprises the following steps:

s1, extracting a user historical behavior log to form basic data, and generating an offline recommendation result by adopting an Ensemble training method;

s2, receiving the real-time behavior log of the user, and generating a near-line recommendation result by adopting a linear fusion method according to the real-time action of the user;

and S3, fusing the offline recommendation result in the step S1 with the online recommendation result in the step S2, and generating a recommendation result by adopting a priority algorithm.

In step S1, the present invention extracts a long-term and massive user history behavior log to form basic data, and then generates an offline recommendation result by using an Ensemble training method for the basic data. Taking the optimized click rate as an example, the invention can form basic data by using item displayed by the user and history behavior logs such as whether to click or not.

The classifiers in the Ensemble training method comprise an L1 layer classifier and an L2 layer classifier; the L1 layer classifier is a basic classifier and can be classified by using basic algorithms such as collaborative filtering, matrix decomposition, contentbase and the like; the L2 layer classifier forms the classification result of the L1 layer classifier into a feature vector based on the L1 layer classifier, and forms the input of the L2 layer classifier (such as GBDT) after combining some other features.

The invention extracts the historical behavior log of the user to form basic data, and generates an offline recommendation result by adopting an Ensemble training method, which specifically comprises the following steps:

s11, taking the basic data as a training sample and dividing the training sample into a Trainpig part and a Test Pig part according to a ratio of 1: 1;

s12, extracting features of the Train pig part in the step S11 to form feature vectors, and training an L1-layer classification model;

s13, predicting the Test Pig part in the step S11 by adopting the L1 layer classification model in the step S12, and generating a L1 layer prediction result;

s14, generating an input feature vector of an L2 layer according to the L1 layer prediction result in the step S13, and training an L2 layer classification model for the Test Pig part;

s15, retraining the L1-layer classification model for all training samples;

and S16, extracting features of the sample to be tested, generating two prediction results by respectively adopting the L1-layer classification model in the step S12 and the L1-layer classification model retrained in the step S15, and generating an offline recommendation result according to the two prediction results by adopting the L2-layer classification model in the step S14.

According to the method, the division proportion of training samples is preset according to different requirements, and the samples to be tested are analyzed according to the L1-layer classification model and the L2-layer classification model obtained through training, so that an offline recommendation result is obtained.

In step S2, the present invention receives a behavior log of the user in real time, where the behavior log includes a presentation log and a click log. The presentation log comprises presented items, algorithms adopted when the items are recommended, recommended positions and item corresponding weights.

According to the real-time action behavior of the user, a linear fusion method which is adjusted by clicking feedback is adopted to generate a near-line recommendation result, namely, the weight of a strategy corresponding to the item recommended is updated according to the real-time action behavior of the user, and the recommendation result is recalculated, wherein the method comprises the following two conditions:

if the real-time action behavior of the user is the presentation log, reducing the weight of the recommended strategy corresponding to the item, and updating the calculation formula of the weight of the recommended strategy corresponding to the item is represented as:

wherein, β'_kFor updated weights, β_kAs the original weight, Ctr_iTo show the average click rate for position i, Score_kScore for algorithm K to this item_itemAnd lambda is the decay constant of the position click rate and the decay constant of the algorithm click rate, wherein lambda is the total score of the item.

If the real-time action behavior of the user is a click log, increasing the weight of the recommended strategy corresponding to the item, and updating the calculation formula of the weight of the recommended strategy corresponding to the item is represented as:

wherein,is the click decay constant.

According to the method, the weighted linear model corresponding to the user is generated according to the calculation formula for updating the weight of the recommended item corresponding strategy, so that the corresponding recommendation result is recalculated according to the weighted linear model.

In step S3, in order to respond more quickly, the present invention uses a priority algorithm to fuse the offline recommendation result in step S1 with the nearline recommendation result in step S2, and generates a final recommendation result. The priority algorithm is a common technical means for those skilled in the art, and the detailed description of the present invention is omitted.

In order to further explain the multidimensional interconnection recommendation method in detail, the invention also provides a multidimensional interconnection recommendation system, which comprises the following steps:

the off-line subsystem is used for establishing a training sample according to the historical behavior log of the user, generating an off-line recommendation result and sending the off-line recommendation result to the on-line subsystem;

the online system comprises an online subsystem and an online subsystem, wherein the online subsystem is used for receiving a real-time behavior log of a user, generating an online recommendation result according to the real-time behavior log of the user and sending the online recommendation result to the online subsystem;

and the online subsystem is used for receiving and fusing the offline recommendation result generated by the offline subsystem and the online recommendation result generated by the online subsystem, and feeding the fused recommendation result back to the user.

The off-line subsystem is arranged at an off-line end, and generates an off-line recommendation result and sends the off-line recommendation result to the on-line subsystem by analyzing commodities, users and a recommendation algorithm by adopting an Ensemble training method according to basic data such as a user historical behavior log stored in a database.

The online system is arranged at the server side, receives the behavior log of the user in real time, generates an online recommendation result by adopting a linear fusion method for adjusting through click feedback according to the real-time action behavior of the user, and sends the online recommendation result to the online system.

The online subsystem of the invention is directly oriented to the user side, and provides high-performance and high-availability recommendation service; the online subsystem receives the offline recommendation result generated by the offline subsystem and the online recommendation result generated by the online subsystem, and generates a recommendation result by adopting a priority algorithm, so that the request pressure of the online subsystem is effectively reduced, and the recommendation result is returned in a short time.

It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.

Claims (10)

1. A multi-dimensional interconnection recommendation method is characterized by comprising the following steps:

s1, extracting a user historical behavior log to form basic data, and generating an offline recommendation result by adopting an Ensemble training method;

s2, receiving the real-time behavior log of the user, and generating a near-line recommendation result by adopting a linear fusion method according to the real-time action of the user;

and S3, fusing the offline recommendation result in the step S1 with the online recommendation result in the step S2, and generating a recommendation result by adopting a priority algorithm.

2. The multidimensional interconnection recommendation method of claim 1, wherein the Ensemble training method comprises a L1 level classifier and a L2 level classifier.

3. The multidimensional interconnection recommendation method according to claim 2, wherein the step S1 is to extract a user history behavior log to form basic data, and the step of generating the offline recommendation result by using the Ensemble training method specifically includes the following sub-steps:

s11, taking the basic data as a training sample and dividing the training sample into a Train Pig part and a Test Pig part according to a ratio of 1: 1;

s12, extracting features of the Train pig part in the step S11 to form feature vectors, and training an L1-layer classification model;

s13, predicting the Test Pig part in the step S11 by adopting the L1 layer classification model in the step S12, and generating a L1 layer prediction result;

s14, generating an input feature vector of an L2 layer according to the L1 layer prediction result in the step S13, and training an L2 layer classification model for the Test Pig part;

s15, retraining the L1-layer classification model for all training samples;

and S16, extracting features of the sample to be tested, generating two prediction results by respectively adopting the L1-layer classification model in the step S12 and the L1-layer classification model retrained in the step S15, and generating an offline recommendation result according to the two prediction results by adopting the L2-layer classification model in the step S14.

4. The multidimensional interconnection recommendation method of claim 3, wherein the behavior log comprises a presentation log and a click log.

5. The multi-dimensional interconnection recommendation method of claim 4, wherein the presentation log comprises presented items, algorithms adopted when the items are recommended, recommendation positions and item corresponding weights.

6. The multi-dimensional interconnection recommendation method according to claim 4, wherein the step S2 of generating the nearline recommendation result by using a linear fusion method according to the real-time action of the user specifically comprises: and updating the weight of the recommended strategy corresponding to the item according to the real-time action behavior of the user, and recalculating the recommendation result.

7. The multidimensional interconnection recommendation method of claim 5, wherein if the real-time action of the user is presenting a log, the weight of the recommended item corresponding strategy is reduced, and the calculation formula is represented as:

${\mathrm{\β}}_k^{\′}={\mathrm{\β}}_k*(1-\mathrm{\λ}*{\mathrm{Ctr}}_i)*(1-\mathrm{\ϵ}*\frac{{\mathrm{Score}}_k}{{\mathrm{Score}}_{item}})$

wherein, β'_kFor updated weights, β_kAs the original weight, Ctr_iTo show the average click rate for position i, Score_kScore for algorithm K to this item_itemAnd lambda is the decay constant of the position click rate and the decay constant of the algorithm click rate, wherein lambda is the total score of the item.

8. The multidimensional interconnection recommendation method of claim 6, wherein if the real-time action behavior of the user is click log, the weight of the recommended strategy corresponding to item is increased, and the calculation formula is represented as:

${\mathrm{\β}}_k^{\′}={\mathrm{\β}}_k*(1+\∂*\frac{{\mathrm{Score}}_k}{{\mathrm{Score}}_{item}})$

wherein,is the click decay constant.

9. A multidimensional interconnection recommendation system, comprising:

the off-line subsystem is used for establishing a training sample according to the historical behavior log of the user, generating an off-line recommendation result and sending the off-line recommendation result to the on-line subsystem;

the online system comprises an online subsystem and an online subsystem, wherein the online subsystem is used for receiving a real-time behavior log of a user, generating an online recommendation result according to the real-time behavior log of the user and sending the online recommendation result to the online subsystem;

and the online subsystem is used for receiving and fusing the offline recommendation result generated by the offline subsystem and the online recommendation result generated by the online subsystem, and feeding the fused recommendation result back to the user.

10. The multidimensional interconnection recommendation system of claim 9, wherein the near-line subsystem is disposed at a server side, and the on-line subsystem faces a client side.