CN110930223A - Recommendation recall method, device and storage medium based on field-aware factorization machine - Google Patents
Recommendation recall method, device and storage medium based on field-aware factorization machine Download PDFInfo
- Publication number
- CN110930223A CN110930223A CN201911145480.XA CN201911145480A CN110930223A CN 110930223 A CN110930223 A CN 110930223A CN 201911145480 A CN201911145480 A CN 201911145480A CN 110930223 A CN110930223 A CN 110930223A
- Authority
- CN
- China
- Prior art keywords
- user
- domain
- item
- article
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0623—Item investigation
- G06Q30/0625—Directed, with specific intent or strategy
- G06Q30/0629—Directed, with specific intent or strategy for generating comparisons
Landscapes
- Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Development Economics (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a recommendation recall method, a device and a storage medium based on a field perception factorization machine, wherein the method comprises the following steps: calculating a domain vector corresponding to the item side feature domain of each item through the item side feature domain of each item in the item library; calculating a domain vector of a user side feature domain of a user to be recommended through the user side feature domain of the user to be recommended; calculating Si and Su; and calculating the scores of the items in the item library through the domain vector of the user to be recommended, the domain vector of the items in the item library, Si and Su. Compared with the prior art, the recommendation recall method based on the field perception factor decomposition machine disclosed by the invention has the advantages that when the FFM algorithm is used in the recommendation recall step, all commodities are quickly and accurately graded and sorted according to the user to be recommended, so that the commodity or information which is more in line with the mind is recommended to the user; meanwhile, one model is used for replacing a plurality of recall models, the recall flow is simplified, and the multi-channel recall hyper-parameter setting is omitted.
Description
Technical Field
The invention belongs to the technical field of computers, and particularly relates to a recommendation recall method, recommendation recall equipment and a storage medium based on a field perception factorization machine.
Background
With the development of Information technology and the internet, people gradually move from the Information-poor era to the Information-Overload (Information-Overload) era. In this age, as information (including merchandise) consumers, it is very difficult to find information of interest to themselves from a large amount of information. The recommendation System (Recommended System) is an important tool for solving the above problems. The task of the recommendation system is to help the user find information that is valuable to him.
The framework of the current mainstream recommendation system is mainly divided into two parts: the first part is recall and the second part is reorder. The recall mainly has the function of primarily screening out commodities which are interested by a target user from a large number of objects to be selected. Because the recall is aimed at prescreening, the recall algorithm is generally low in computational complexity and meets the requirement of rapid screening. Since the behavior of the user is diverse, multiple recall loops may be designed from different perspectives. Taking the text recommendation as an example, the text has different features such as a topic, an entity, a tag, and the like, and 3 recalls can be designed according to the three features of the topic of interest, the entity of interest, and the tag of interest of the user. And then, the reordering is to order the commodities recalled in the plurality of recalling circuits according to the scores, select the commodity with higher score and recommend the commodity to the user.
However, this recommendation system uses a simple algorithm as a recall algorithm, and only linear information about features can be captured at the time of recall, but non-linear information, particularly cross-feature information, cannot be captured, and therefore accuracy is not high. In addition, the recommendation system needs to design multiple recalls, and each recall has one hyper-parameter to be debugged.
Disclosure of Invention
The invention aims to provide a recommendation recall method, a recommendation recall device and a recommendation recall storage medium based on a field perception factorization machine.
In order to achieve one of the above objects, an embodiment of the present invention provides a recommendation recall method based on a locale perceptual factorizer, the method including:
calculating a domain vector corresponding to the item side feature domain of each item through the item side feature domain of each item in the item library;
calculating a domain vector of a user side feature domain of a user to be recommended through the user side feature domain of the user to be recommended;
calculating a first-order item score Si of the item side characteristic of the item in the item library and a first-order item score Su of the user side characteristic of the user to be recommended;
and calculating the scores of the items in the item library through the domain vector of the user to be recommended, the domain vector of the items in the item library, the Si and the Su, and selecting the items with the scores of N before as recommended recall items.
As a further improvement of an embodiment of the present invention, the "calculating a score of an item in an item library through a domain vector of a user to be recommended, a domain vector of an item in an item library, Si, and Su" specifically includes:
and respectively carrying out vector inner products on all the domain vectors of the user to be recommended and the corresponding domain vectors of the articles in the article library, summing the vector inner products, and adding the corresponding Su and Si to obtain the score of each article relative to the user to be recommended.
As a further improvement of an embodiment of the present invention, the "calculating a score of an item in an item library through a domain vector of a user to be recommended, a domain vector of an item in an item library, Si, and Su" specifically includes:
sequentially splicing the domain vectors, 1 and Si of all article side characteristic domains of each article to obtain the vector of each article;
sequentially splicing the domain vectors, Su and 1 of all user side characteristic domains of a user to be recommended to obtain a retrieval vector of the user to be recommended;
and performing vector inner product on the retrieval vector and the vector of each article in the article library, and selecting the article with the top N as a recommended recall article.
As a further improvement of an embodiment of the present invention, vectors of the items in the item library are stored in a vector database, when an item needs to be recommended to a user to be recommended, the vector database is searched using a search vector of the user to be recommended, a vector inner product of the search vector and the vector of the item is used as a search score, and an item N before the search score is selected as a recommendation recall item.
As a further improvement of an embodiment of the present invention, the "calculating a domain vector corresponding to the article-side feature domain of each article by using the article-side feature domain of each article in the article library" specifically includes:
and calculating r domain vectors of each article side feature domain relative to r user side feature domains through m article side feature domains of each article in the article library, and obtaining r x m domain vectors in total for each article.
As a further improvement of an embodiment of the present invention, the "calculating a domain vector of a user-side feature domain of a user to be recommended by using the user-side feature domain of the user to be recommended" specifically includes:
and calculating m domain vectors of each article side feature domain relative to the m article side feature domains through r user side feature domains of the user to be recommended, and obtaining m x r domain vectors in total.
As a further improvement of an embodiment of the present invention, the method further comprises:
and (4) sorting the user side characteristic and the article side characteristic, and training the FFM to obtain related parameters of the FFM.
As a further improvement of an embodiment of the present invention, the "sorting the user-side features and the article-side features, training the FFM model, and obtaining parameters related to the FFM model" specifically includes:
sorting the article side features and the user side features, dividing the article side features into m feature domains, and dividing the article side features into r feature domains;
collecting behavior records of historical users, training an FFM (fringe field model) model, and obtaining m embedding vectors of each user side feature relative to m feature domains of the article side feature, the weight of a first-order term of the user side feature, r embedding vectors of each article side feature relative to the user side feature and the weight of the first-order term of the article side feature.
In order to achieve one of the above objects, an embodiment of the present invention provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program operable on the processor, and the processor executes the computer program to implement the steps in any one of the above recommended recall method based on a perceptron factorizer.
To achieve one of the above objects, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of any of the above methods for recalling recommendations based on a perceptron factorizer.
Compared with the prior art, the recommendation recall method based on the field perception factor decomposition machine disclosed by the invention has the advantages that when the FFM algorithm is used in the recommendation recall step, all commodities are quickly and accurately graded and sorted according to the user to be recommended, so that the commodity or information which is more in line with the mind is recommended to the user; meanwhile, one model is used for replacing a plurality of recall models, the recall flow is simplified, and the multi-channel recall hyper-parameter setting is omitted.
Drawings
Fig. 1 is a flowchart illustrating a recommendation recall method according to a first embodiment of the present invention.
FIG. 2 is a flowchart illustrating a recommendation recall method based on a FED according to a second embodiment of the present invention.
Detailed Description
The present invention will be described in detail below with reference to specific embodiments shown in the drawings. These embodiments are not intended to limit the present invention, and structural, methodological, or functional changes made by those skilled in the art according to these embodiments are included in the scope of the present invention.
The traditional recommendation system comprises two parts of recall and reordering, and the recall algorithm is generally low in computational complexity and needs to meet the requirement of fast screening, but needs to set multi-way recall. Algorithms used in reordering generally have complicated calculation, but have high accuracy, and can calculate cross information between features, so that the results of multi-way recalls can be sorted, and then the items or information with the highest scores can be selected and recommended to users. In the prior art, the FM algorithm (Factorization Machines)/FFM algorithm (Field-aware Factorization Machines) is generally used in the recommendation system as a reordering algorithm.
The FM algorithm model formula is as follows:
where n represents the number of features of the sample, b _ i can also be written as x _ i, representing the value of the ith feature, w _0, w _ i and w _ ij are model parameters, w _ i represents the weight of the first order term features (subsequently, the weight of all user-side features of the first order term and the weight of all item-side features of the first order term are represented by Wu and Wi), and w _ ij is the weight of the second order cross feature. Write wij over into the form of vector inner product:
substituting it into equation (1) yields:
the third term second order cross feature part of equation (2) (referred to as cross term part for short) can be derived as follows:
the final cross term part can be expressed as formula (3), and as can be seen from formula (3), the previous term in the brace is the backward quantity antipodal addition of all the eigenvalues x multiplied by their corresponding imbedding vectors, and the final vector itself is the vector inner product. The second term in the parenthesis is that the eigenvalue x and the element of its corresponding embedding vector are multiplied after each square, and then all the multiplication results are added.
It should be noted that the embedding vector is also called an embedded vector or a hidden vector, and is a k-dimensional vector trained by the FM model for each feature, and each feature has one embedding vector.
From the above derived results, we consider the features of the sample to be divided into user-side features and item-side features, and then make some conditional restrictions on the cross term part, such as not considering feature cross inside the user-side features and feature cross inside the item-side features, and the above cross term part can continue to be simplified as follows (k is the dimension of the embedding vector):
the feature intersection of the user-side feature and the article-side feature is shown in the formula (4), and it can be seen that the final calculation is to multiply the user-side feature by the embedding vector of the user-side feature, then add the vectors of all the user-side features to obtain a user feature vector, perform the same operation on the article side to obtain an article feature vector, and then perform the inner product on the two vectors. There is no following squaring term compared to equation (3).
According to the above derivation result, the features are divided into a user side (U) and an article side (I), and the two types of features do not consider feature crossing inside each type (in practical application, the influence of the feature crossing inside each type is relatively small, for example, the relationship between the user age and the user gender is not large), so that the cross term part of the FM model can be simplified as the vector inner product of the user side feature vector Vu and the article side feature vector Vi. The FM model further includes a first order term (only the value of the user-side feature or the item-side feature and the weight of the first order term of the corresponding feature need to be vector-inner-multiplied when calculating the score of the first order term) and a constant w _0, which is not considered because the constant w _0 has no influence on the final integration order. When Su represents the user-side first-order score and Si represents the item-side first-order score, the key point in calculating the product score according to the user is the calculation result F of the following formula (5):
F＝Vu*Vi+Su+Si (5)
where Vu ═ Vu1, Vu2, …, Vuk ], Vi ═ Vi1, Vi2, …, Vik, k are dimensions of the eigen embedding vector. In order to be able to directly represent the user-side related content (hereinafter simply referred to as user's vector VuS) and the item-side related content (hereinafter simply referred to as item's vector ViS) by vectors, respectively, a user's vector VuS and an item's vector ViS are constructed:
VuS＝[Vu1,Vu2,…,Vuk,Su,1]
ViS＝[Vi1,Vi2,…,Vik,1,Si]
the commodity score F-VuS ViS is thus calculated from the user.
The above is a simplification process of the FM model, the FFM algorithm introduces a concept of a feature domain on the basis of the FM algorithm, and the FFM model formula is as follows:
each feature in the FM model has an embedding vector. The FFM model introduces the concept of a feature domain, and the FFM attributes features with the same property to the same feature domain, so that in the FFM, each dimension feature x _ i aims at each feature domain fj of other features to learn an embedding vector v _ i, fj. Rewrite w _ ij to the form of vector inner product:
substituting it into equation (6) yields:
in the FM model, the features of the sample are divided into a user-side feature and an article-side feature, and then some conditional restrictions are applied to the cross term part, for example, the feature cross inside the user-side feature and the feature cross inside the article-side feature are not considered, so that the part of the second-order cross feature can be simplified. Similarly, in the FFM model, the features of the sample are also divided into the user-side feature and the article-side feature, and then the feature intersection inside the user-side feature and the feature intersection inside the article-side feature are ignored, and the part of the third-term second-order cross feature of equation (7) (referred to as cross term part for short) may be derived as follows:
the final cross term portion can be expressed as equation (8), where f _ U and f _ I in the equation represent the total number of feature fields of the user-side feature and the item-side feature, respectively, r and m represent the several feature fields of the user-side feature and the item-side feature, respectively, fr and fm represent the total number of features in the r-th feature field of the user-side feature and the m-th feature field of the item-side feature, respectively, and k is the dimension of the embedding vector.
Referring to the parenthesized part of equation (8), the feature fields r and m are regarded as fixed values, that is, the feature field r of the user-side feature is directed to the feature field m of the article-side feature, and the feature value in the feature field is multiplied by the corresponding embedding vector and added in place to obtain a total k-dimensional vector of one feature field relative to another feature field, where the total k-dimensional vector is also called a field vector, abbreviated as E, as follows:
then, E of the two feature domains is subjected to vector inner product, i.e., EU _ rm · EI _ mr, EU represents the domain vector of the user-side feature, EI represents the domain vector of the article-side feature, EU _ rm represents the domain vector of the r-th feature domain of the user-side feature relative to the m-th feature domain of the article-side feature, EI _ mr represents the domain vector of the m-th feature domain of the article-side feature relative to the r-th feature domain of the user-side feature, and "·" represents the vector inner product.
Please refer to the formula (8), the whole cross term part is equivalent to that the domain vector of each feature domain of the user-side feature and the domain vector of each feature domain of the article-side feature domain are vector-inner-product and summed. Specifically, the user-side features total r feature fields, the item-side features total m feature fields, for the first feature field of the user-side features, which has m field vectors including EU _11, EU _12 to EU _1m with respect to the m feature fields of the item-side features, and the m feature fields of the item-side features have field vectors EI _11, EI _21 to EI _ m1 with respect to the first feature field of the user-side features, respectively.
By analogy, the r-th feature domain of the user-side feature has domain vectors EU _ r1, EU _ r2 to EU _ rm with respect to the m feature domains of the item-side feature, and the m feature domains of the item-side feature have domain vectors EI _1r, EI _2r to EI _ mr with respect to the r-th feature domain of the user-side feature, respectively.
The user side features have r × m domain vectors in total, the article side features also have m × r domain vectors, and finally all the domain vectors EU _ rm of the user side features and the corresponding domain vectors EI _ mr of the article side features are subjected to vector inner product and summed, that is, formula (8) can be simplified as follows:
according to the above derivation result, the features are divided into a user side (U) and an article side (I), and the two types of features do not consider feature crossing inside each type (in practical application, the influence of the feature crossing inside each type is relatively small, for example, the relationship between the age of the user and the gender of the user is not large), so that the cross term part of the FFM model can be simplified into that each domain vector of each feature domain of the features on the user side and the domain vector of the feature domain of the corresponding article side features are subjected to vector inner product and summation.
The FFM model further includes a first-order term (only the value of the user-side feature or the item-side feature and the weight of the first-order term of the corresponding feature need to be vector-inner-multiplied when calculating the score of the first-order term) and a constant w _0, and the constant w _0 does not have an influence on the final integration ordering, so that the constant w _0 is not considered. When Su represents the user-side first-order score and Si represents the item-side first-order score, the key point in calculating the product score from the user is the calculation result F of the following formula (10):
as in FM, in order to directly represent the user-side related content (hereinafter simply referred to as user vector VuS) and the item-side related content (hereinafter simply referred to as item vector ViS), respectively, a user vector VuS and an item vector ViS are constructed:
VuS＝[EU_11,EU_12,…,EU_1m,…,EU_r1,EU_r2,…,EU_rm,Su,1]
ViS＝[EI_11,EI_21,…,EI_m1,…,EI_1r,EI_2r,…,EI_mr,1,Si]
the commodity score F-VuS ViS is thus calculated from the user.
From the above derivation, it can be seen that the calculation process is greatly simplified when the FM and FFM model formulas are used to calculate the score of an article. Therefore, the FM and FFM algorithms can be used in recalling of the recommendation system, and when the FM and FFM algorithms are used in the recommendation recall, the cross features of the user side features and the article side features can be captured, so that all commodities can be quickly and accurately sequenced according to the user, and commodities or information which are more in line with the mind can be recommended to the user; meanwhile, the method replaces the step of multi-way recall, simplifies the recall flow and saves the super-parameter setting of each recall.
In addition, it should be noted that, since the simplified FM/FFM model omits the internal cross feature of the user-side feature and the internal cross feature of the item-side feature, and therefore the accuracy is lower than that of the original FM/FFM model, it is preferable to re-order the recalled items using a model with higher accuracy (such as the original FM/FFM model) after the recall, in this process, the number of recalls is preferably greater than the number recommended to the user after the re-ordering, that is, if it is necessary to recommend 10 items to the user, 100 items are recalled, and then the 100 items are re-ordered, and the top 10 items are selected and recommended to the user. Of course, if the accuracy requirement is not very high, the recalled item may be directly recommended to the user, and the reordering step may be omitted.
According to the derivation result, a first embodiment of the present invention provides a recommendation recall method, which may be implemented by using an FM model or an FFM model, as shown in fig. 1, and the method includes:
step S110: from the item-side features, a vector is calculated for each item in the item library.
When the FM model is adopted, the step S110 specifically includes:
step S111: and training an FM model to obtain an embedding vector of the article side characteristic.
The article-side features and the user-side features are first sorted, for example, the article-side features may include an article ID, an article tag, an article price, and the like, and the user-side features may include a user ID, a user gender, a user age, and the like.
Then, collecting records of historical user clicks or browses of articles, training an FM model, and obtaining embedding vectors of all the features and weights of first-order items, wherein the embedding vectors of all the user-side features, the embedding vectors of all the article-side features, the weights Wu of the first-order items of the user-side features and the weights Wi of the first-order items of the article-side features are specifically included.
Step S112: and adding vector pairs obtained by multiplying the value of each article side characteristic of each article in the article library by the corresponding embedding vector, and obtaining a characteristic vector Vi for each article.
The method comprises the steps of storing a plurality of articles in inventory, calculating a feature vector of one article, namely multiplying a value of each article side feature of the article by a corresponding embedding vector to obtain a plurality of k-dimensional vectors, and performing opposite addition on the k-dimensional vectors to obtain a feature vector Vi of the article.
As a simple example, assume that there are 3 item-side features, and their corresponding embedding vectors are A, B, C, where a ═ a1, a2, …, ak ], B ═ B1, B2, …, bk, and C ═ C1, C2, …, ck. If the item-side feature values of an item are 1, and 1, respectively, the feature vector Vi of the item is [ a1+ b1+ c1, a2+ b2+ c2, …, ak + bk + ck ].
Step S113: a first order term score Si is calculated for the item-side feature of each item in the item library.
In the previous training of the FM model, the weight Wi of the first-order term of the item-side feature is obtained, and the value of each item-side feature of the items in the item library is weighted and summed with the Wi to obtain the item-side first-order term score Si of each item.
Assuming that there are m item-side features, then Wi ═ Wi1, Wi2, …, Wim, the value of the item-side feature for one item is [ x1, x2, …, xm ], then Si ═ Wi1 x1+ Wi2 x2+ … + Wim x m.
Step S114: and sequentially splicing the Vi, the 1 and the Si to obtain the vector of the article.
For the sake of storage convenience and fast calculation, the content related to the item side, i.e., the vector ViS of the item, is directly represented by a vector, and assuming Vi ═ Vi1, Vi2, …, Vik ], the ViS ═ Vi1, Vi2, …, Vik, Si, 1.
When the FFM model is adopted, the step S110 specifically includes:
step S111': and sorting the article side characteristic and the user side characteristic, and training the FFM model.
Sorting the article side features and the user side features, dividing the article side features into m feature domains, and dividing the article side features into r feature domains;
collecting behavior records (browsing history or purchase history and the like) of historical users, training an FFM (fringe field model) to obtain m embedding vectors of m feature fields of each user-side feature relative to the article-side feature, the weight of a first-order item of the user-side feature, r embedding vectors of each article-side feature relative to the user-side feature and the weight of a first-order item of the article-side feature.
Step S112': from the item-side features, a vector is calculated for each item in the item library.
R domain vectors of each feature domain of the item-side features of each item in the item library relative to r feature domains of the user-side features are calculated, resulting in r m domain vectors for each item in total, EU _11, EU _12, …, EU _1m, …, EU _ r1, EU _ r2, …, EU _ rm.
And calculating the first-order item score Si of the item-side characteristic of each item in the item library, which is the same as the calculation step of the FM model.
And sequentially splicing the r × m domain vectors, 1 and Si to obtain a vector Vis of the article [ EI _11, EI _21, …, EI _ m1, …, EI _1r, EI _2r, …, EI _ mr,1, Si ].
In the above step, since one item-side feature is trained to obtain one embedding vector with respect to each feature domain of the user-side feature, calculating a domain vector EI _ mr of one feature domain m of the item-side feature of one item with respect to one feature domain r of the user-side feature specifically includes:
and adding vector alignment obtained by multiplying the feature value of each feature of the feature domain m of the article-side feature of the article by the embedding vector of the feature domain r corresponding to the user-side feature to obtain a domain vector EI _ mr of the feature domain m of the article-side feature relative to the feature domain r of the user-side feature.
As a simple example, assuming that the feature field m of the article-side features includes 3 features, the values of these three features of the article are a, b and c, respectively, and the embedding vectors of the 3 features of the feature field m of the article-side features with respect to the feature field r of the user-side features are e1, e2 and e3, respectively, then the EI _ mr is a e1+ b e2+ c e 3.
Through the above step S110, a vector of each item in the item library may be calculated, and preferably, the items in the item library are stored in a vector form.
Under the condition that the characteristics of the articles in the article library are not changed, the vector of each article only needs to be calculated once, and the method can be used all the time, and is simple and convenient.
It should be noted that the items in the item library of the present invention are not limited to the actual sold goods, but may also represent any other recommended information, such as news, video, music, etc.
Step S120: and calculating the retrieval vector of the user to be recommended according to the user side characteristics.
When the FM model is adopted, the step S120 includes:
computing a retrieval vector with recommended users is similar to computing a vector of items. The foregoing training FM model already obtains the embedding vector of the user-side feature, and adds the vector pair obtained by multiplying the value of the user-side feature of the user to be recommended by the corresponding embedding vector, so as to obtain the feature vector Vu of the user to be recommended [ Vu1, Vu2, …, Vuk ].
Calculating a first-order item score Su of the user-side characteristics of the user to be recommended: assuming that there are r user-side features, Wu ═ Wu1, Wu2, …, Wur, the user-side feature with recommended user has a value of [ x1, x2, …, xr ], Su ═ Wu1 x1+ Wu2 x2+ … + Wur xr.
And sequentially splicing the Vu, Su and 1 to obtain a retrieval vector VuS of the user to be recommended, wherein the VuS is [ Vu1, Vu2, …, Vuk, Su,1 ].
In a preferred embodiment, the user-side features are divided into fixed features and varying features; the fixed features may include features that generally do not change after the user logs in the website, such as user id, user gender, and user age; the changed features may include a time the user visits, a place the user visits, an interest tag of the user, a browsing history of the user, etc., and may change over time after the user logs in to the website.
Therefore, the feature vector Vu of the user to be recommended includes the fixed feature vector Vuf and the varied feature vector Vuv, and the first-order item score Su of the user-side feature of the user to be recommended includes the first-order item score Suf of the fixed feature and the first-order item score Suv of the varied feature.
Since the user needs to recommend the user regularly after being online, when the retrieval vector VuS of the user to be recommended is calculated, only Vuv and Suv can be calculated for subsequent recommendations except for the calculation of Vuf and Suf for the first recommendation, thereby speeding up the calculation.
When the FFM model is adopted, the step S120 specifically includes:
calculating m domain vectors of m feature domains relative to the item-side features of each feature domain of the user-side features of the user to be recommended, and obtaining m r domain vectors in total, wherein the m r domain vectors are EU _11, EU _12, …, EU _1m, …, EU _ r1, EU _ r2, … and EU _ rm.
And calculating the first-order item score Su of the user-side characteristic of the user to be recommended, wherein the step is the same as the step of calculating the FM model.
And sequentially splicing the m × r domain vectors, Su and 1 to obtain a retrieval vector VuS of the user to be recommended [ EU _11, EU _12, …, EU _1m, …, EU _ r1, EU _ r2, …, EU _ rm, Su,1 ].
In the above step, calculating a domain vector EU _ rm of a feature domain r of the user-side feature relative to a feature domain m of the article-side feature specifically includes:
and adding vector alignment obtained by multiplying the feature values of the feature domain r of the user-side feature by the embedding vector of the feature domain m corresponding to the article-side feature to obtain a domain vector EU _ rm of the feature domain r of the user-side feature relative to the feature domain m of the article-side feature.
It should be noted that, as in the FM model, the user-side feature preferably includes a fixed feature and a variable feature. The user-side feature thus comprises one or more fixed feature fields and one or more varying feature fields. When the domain vector is calculated, the domain vector of the fixed characteristic domain only needs to be calculated once, and then the domain vector of the variable characteristic domain is calculated in real time, so that the speed of recommendation recall is increased.
Step S130: and performing vector inner product on the retrieval vector and the vector of each article in the article library, and selecting the article with the top N as a recommended recall article.
Using the FM model, the item score F VuS ViS Vi + Su + Si. And selecting the items with the top N as recommended recall items, and directly recommending the recommended recall items to the user, or recommending the recommended recall items to the user after reordering the recommended recall items.
And (3) adopting an FFM (fringe field model), when a user to be recommended makes a recommendation recall, calculating the score of each item according to a formula (10), and taking the item N before the score as a recommendation recall item.
In a preferred embodiment, in order to speed up the recommendation or calculation, vectors of articles in the article library are stored in a vector database, when articles need to be recommended to a user to be recommended, the vector database is searched by using the retrieval vector of the user to be recommended, the vector inner product of the retrieval vector and the vector of the articles is used as a retrieval score, and the articles N before the retrieval score are selected as recommendation recall articles.
In a specific embodiment, a user registers basic information in a commodity library (an item library), selects commodities (which can be browsed or purchased) in the commodity library after the commodities are online, trains an FM model according to historical user behaviors and sorted item-side features and user-side features (the user-side features include fixed features and variable features), and obtains embedding vectors and weights of first-order terms of the features, specifically including the embedding vectors of the user-side features, the embedding vectors of the item-side features, the weights Wu of the first-order terms of the user-side features, and the weights Wi of the first-order terms of the item-side features. These data are stored so that when the goods are subsequently added or modified, vectors for the goods are calculated.
And calculating the vector of each commodity in the commodity library according to the trained FM model, and storing the vector in a vector database. After a certain user is on line, according to the characteristics of the user, such as gender, age, on-line location, browsing history (if any), and the like, calculating a retrieval vector of the user, performing vector inner product on the retrieval vector and the vectors of all commodities in the commodity library, selecting N commodities before scoring as recommended recall commodities, reordering the recommended recall commodities, and recommending to the user.
In another specific embodiment, a user registers basic information in a commodity library (an article library), and selects commodities (which can be browsed or purchased) in the commodity library after online, according to the behaviors of historical users and sorted article side features and user side features (the user side features comprise fixed features and variable features), the article side features are divided into m feature fields, the article side features are divided into r feature fields, an FFM model is trained, and m embedding vectors of each user side feature relative to the m feature fields of the article side features, the weight of a first order of the user side features, and the weight of r embedding vectors of each article side feature relative to the user side features and the weight of a first order of the article side features are obtained. These data are stored so that when the goods are subsequently added or modified, vectors for the goods are calculated.
And calculating the vector of each commodity in the commodity library according to the trained FFM model, and storing the vector in a vector database. After a certain user is on line, according to the characteristics of the user, such as gender, age, on-line location, browsing history (if any), and the like, calculating a retrieval vector of the user, performing vector inner product on the retrieval vector and the vectors of all commodities in the commodity library, selecting N commodities before scoring as recommended recall commodities, reordering the recommended recall commodities, and recommending to the user. As the user continuously browses, the value of the change characteristic of the user continuously changes, and thus the recommended product continuously changes. During calculation, the user-side feature can be divided into two parts for calculation, the value of the change feature is calculated in real time, and the value of the fixed feature calculated for the first time is saved.
The invention further provides an electronic device, which includes a memory and a processor, where the memory stores a computer program operable on the processor, and the processor implements any one of the steps of the recommended recall method when executing the program, that is, implements the steps in any one of the technical solutions of the recommended recall method.
The present invention also provides a computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, implements any one of the steps of the above-mentioned recommendation recall method, that is, implements the steps of any one of the above-mentioned recommendation recall method.
As shown in fig. 2, the second embodiment of the present invention further provides a recommendation recall method based on a perceptual factorization machine, where the method includes:
step S210: and calculating a domain vector corresponding to the item-side characteristic domain of each item through the item-side characteristic domain of each item in the item library.
The step S210 specifically includes: and calculating r domain vectors of each article side feature domain relative to r user side feature domains through m article side feature domains of each article in the article library, and obtaining r x m domain vectors in total for each article. Note that, herein, the item-side feature field refers to a feature field of an item-side feature, and the user-side feature field refers to a feature field of a user-side feature.
Step S220: and calculating the domain vector of the user side characteristic domain of the user to be recommended according to the user side characteristic domain of the user to be recommended.
The step S220 specifically includes: and calculating m domain vectors of each article side feature domain relative to the m article side feature domains through r user side feature domains of the user to be recommended, and obtaining m x r domain vectors in total.
Step S230: and calculating the first-order item score Si of the item-side characteristic of the item in the item library and the first-order item score Su of the user-side characteristic of the user to be recommended.
Step S240: and calculating the scores of the items in the item library through the domain vector of the user to be recommended, the domain vectors of the items in the item library, Si and Su, and selecting the items N before the scores as the recommended recall items.
The step S240 specifically includes:
and respectively carrying out vector inner products on all the domain vectors of the user to be recommended and the corresponding domain vectors of the articles in the article library, summing the vector inner products, and adding the corresponding Su and Si to obtain the score of each article relative to the user to be recommended.
In order to increase the calculation speed, in a preferred embodiment, the step S240 includes:
sequentially splicing the domain vectors, 1 and Si of all article side characteristic domains of each article to obtain the vector of each article;
sequentially splicing the domain vectors, Su and 1 of all user side characteristic domains of a user to be recommended to obtain a retrieval vector of the user to be recommended;
and performing vector inner product on the retrieval vector and the vector of each article in the article library, and selecting the article with the top N as a recommended recall article.
In order to further increase the calculation speed, it is preferable that vectors of the articles in the article library are stored in a vector database, when an article needs to be recommended to a user to be recommended, the vector database is searched by using a search vector of the user to be recommended, a vector inner product of the search vector and the vector of the article is used as a search score, and the article N before the search score is selected as a recommendation recall article.
In a specific embodiment, the method further comprises:
the user side characteristic and the article side characteristic are sorted, the FFM model is trained, and parameters related to the FFM model are obtained, namely:
sorting the article side features and the user side features, dividing the article side features into m feature domains, and dividing the article side features into r feature domains; collecting behavior records of historical users, training an FFM (fringe field model) model, and obtaining m embedding vectors of each user side feature relative to m feature domains of the article side feature, the weight of a first-order term of the user side feature, r embedding vectors of each article side feature relative to the user side feature and the weight of the first-order term of the article side feature.
Preferably, calculating a domain vector of a feature domain of the item-side features of an item relative to a feature domain of the user-side features specifically comprises:
and adding vector alignment obtained by multiplying the value of each feature of the feature domain of the article-side feature of the article by the embedding vector of the feature domain corresponding to the user-side feature to obtain a domain vector of the feature domain of the article-side feature of the article relative to the feature domain of the user-side feature.
Preferably, the calculating a domain vector of a feature domain of the user-side features of the user to be recommended relative to a feature domain of the item-side features specifically includes:
and adding the vector counterpoint obtained by multiplying the value of each feature of the feature domain of the user side feature of the user to be recommended by the embedding vector of the feature domain of the corresponding article side feature to obtain the domain vector of the feature domain of the user side feature of the user to be recommended relative to the feature domain of the article side feature.
The invention further provides an electronic device, which includes a memory and a processor, where the memory stores a computer program operable on the processor, and the processor implements any one of the steps of the recommended recall method based on the bsp when executing the program, that is, implements the steps of any one of the technical solutions of the recommended recall method based on the bsp.
The present invention further provides a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements any one of the steps in the above recommended recall method based on a perceptual factorizer, that is, implements any one of the steps in the above recommended recall method based on a perceptual factorizer.
It should be understood that although the present description refers to embodiments, not every embodiment contains only a single technical solution, and such description is for clarity only, and those skilled in the art should make the description as a whole, and the technical solutions in the embodiments can also be combined appropriately to form other embodiments understood by those skilled in the art.
The above-listed detailed description is only a specific description of a possible embodiment of the present invention, and they are not intended to limit the scope of the present invention, and equivalent embodiments or modifications made without departing from the technical spirit of the present invention should be included in the scope of the present invention.
Claims (10)
1. A recommendation recall method based on a farm aware factorizer, the method comprising:
calculating a domain vector corresponding to the item side feature domain of each item through the item side feature domain of each item in the item library;
calculating a domain vector of a user side feature domain of a user to be recommended through the user side feature domain of the user to be recommended;
calculating a first-order item score Si of the item side characteristic of the item in the item library and a first-order item score Su of the user side characteristic of the user to be recommended;
and calculating the scores of the items in the item library through the domain vector of the user to be recommended, the domain vector of the items in the item library, the Si and the Su, and selecting the items with the scores of N before as recommended recall items.
2. The recommendation recall method based on the farm-aware factorization machine according to claim 1, wherein the "calculating the score of the item in the item library through the domain vector of the user to be recommended, the domain vectors of the items in the item library, Si and Su" specifically comprises:
and respectively carrying out vector inner products on all the domain vectors of the user to be recommended and the corresponding domain vectors of the articles in the article library, summing the vector inner products, and adding the corresponding Su and Si to obtain the score of each article relative to the user to be recommended.
3. The recommendation recall method based on the farm-aware factorization machine according to claim 1, wherein the "calculating the score of the item in the item library through the domain vector of the user to be recommended, the domain vectors of the items in the item library, Si and Su" specifically comprises:
sequentially splicing the domain vectors, 1 and Si of all article side characteristic domains of each article to obtain the vector of each article;
sequentially splicing the domain vectors, Su and 1 of all user side characteristic domains of a user to be recommended to obtain a retrieval vector of the user to be recommended;
and performing vector inner product on the retrieval vector and the vector of each article in the article library, and selecting the article with the top N as a recommended recall article.
4. The farm-aware factorization machine-based referral recall method of claim 3, wherein:
and storing the vectors of the articles in the article library in a vector database, searching the vector database by using the retrieval vector of the user to be recommended when the articles need to be recommended to the user to be recommended, using the vector inner product of the retrieval vector and the vector of the articles as a retrieval score, and selecting the articles N before the retrieval score as recommended recall articles.
5. The recommendation recall method according to claim 1, wherein the calculating a domain vector corresponding to the item-side feature domain of each item from the item-side feature domain of each item in the item library specifically comprises:
and calculating r domain vectors of each article side feature domain relative to r user side feature domains through m article side feature domains of each article in the article library, and obtaining r x m domain vectors in total for each article.
6. The recommendation recall method based on the FED of claim 1, wherein the calculating the domain vector of the user-side feature domain of the user to be recommended through the user-side feature domain of the user to be recommended specifically comprises:
and calculating m domain vectors of each article side feature domain relative to the m article side feature domains through r user side feature domains of the user to be recommended, and obtaining m x r domain vectors in total.
7. The farm-aware factorization machine based recommendation recall method of claim 1, further comprising:
and (4) sorting the user side characteristic and the article side characteristic, and training the FFM to obtain related parameters of the FFM.
8. The recommendation recall method based on the perceptual factorization machine of claim 7, wherein the "sorting the user-side features and the article-side features and training the FFM model to obtain the parameters related to the FFM model" specifically comprises:
sorting the article side features and the user side features, dividing the article side features into m feature domains, and dividing the article side features into r feature domains;
collecting behavior records of historical users, training an FFM (fringe field model) model, and obtaining m embedding vectors of each user side feature relative to m feature domains of the article side feature, the weight of a first-order term of the user side feature, r embedding vectors of each article side feature relative to the user side feature and the weight of the first-order term of the article side feature.
9. An electronic device comprising a memory and a processor, the memory storing a computer program operable on the processor, wherein the processor when executing the program performs the steps in the farm perception factor decomposition-based recommendation recall method of any one of claims 1-8.
10. A computer-readable storage medium, having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for recommended recall based on a perceptual factorization machine of any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911145480.XA CN110930223A (en) | 2019-11-21 | 2019-11-21 | Recommendation recall method, device and storage medium based on field-aware factorization machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911145480.XA CN110930223A (en) | 2019-11-21 | 2019-11-21 | Recommendation recall method, device and storage medium based on field-aware factorization machine |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110930223A true CN110930223A (en) | 2020-03-27 |
Family
ID=69850520
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911145480.XA Withdrawn CN110930223A (en) | 2019-11-21 | 2019-11-21 | Recommendation recall method, device and storage medium based on field-aware factorization machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110930223A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112817442A (en) * | 2021-01-20 | 2021-05-18 | 北京理工大学 | Situation information classification recommendation system and method under multi-task condition based on FFM |
| CN112949703A (en) * | 2021-02-23 | 2021-06-11 | 广州市百果园信息技术有限公司 | Video recall method and device and electronic equipment |
| CN112984919A (en) * | 2021-05-12 | 2021-06-18 | 创新奇智(北京)科技有限公司 | Refrigerating system energy efficiency optimization method and device, electronic equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180020250A1 (en) * | 2015-09-08 | 2018-01-18 | Tencent Technology (Shenzhen) Company Limited | Recommendation information pushing method, server, and storage medium |
| CN108830680A (en) * | 2018-05-30 | 2018-11-16 | 山东大学 | Personalized recommendation method, system and storage medium based on discrete disassembler |
| CN110162693A (en) * | 2019-03-04 | 2019-08-23 | 腾讯科技（深圳）有限公司 | A kind of method and server of information recommendation |
| CN110457581A (en) * | 2019-08-02 | 2019-11-15 | 达而观信息科技（上海）有限公司 | A kind of information recommended method, device, electronic equipment and storage medium |
-
2019
- 2019-11-21 CN CN201911145480.XA patent/CN110930223A/en not_active Withdrawn
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180020250A1 (en) * | 2015-09-08 | 2018-01-18 | Tencent Technology (Shenzhen) Company Limited | Recommendation information pushing method, server, and storage medium |
| CN108830680A (en) * | 2018-05-30 | 2018-11-16 | 山东大学 | Personalized recommendation method, system and storage medium based on discrete disassembler |
| CN110162693A (en) * | 2019-03-04 | 2019-08-23 | 腾讯科技（深圳）有限公司 | A kind of method and server of information recommendation |
| CN110457581A (en) * | 2019-08-02 | 2019-11-15 | 达而观信息科技（上海）有限公司 | A kind of information recommended method, device, electronic equipment and storage medium |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112817442A (en) * | 2021-01-20 | 2021-05-18 | 北京理工大学 | Situation information classification recommendation system and method under multi-task condition based on FFM |
| CN112817442B (en) * | 2021-01-20 | 2023-07-25 | 北京理工大学 | FFM-based situation information classification recommendation system and method under multitasking condition |
| CN112949703A (en) * | 2021-02-23 | 2021-06-11 | 广州市百果园信息技术有限公司 | Video recall method and device and electronic equipment |
| CN112984919A (en) * | 2021-05-12 | 2021-06-18 | 创新奇智(北京)科技有限公司 | Refrigerating system energy efficiency optimization method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10409821B2 (en) | Search result ranking using machine learning | |
| CN108230058B (en) | Product recommendation method and system | |
| CN110866191A (en) | Recommendation recall method, apparatus and storage medium | |
| US9208437B2 (en) | Personalized information pushing method and device | |
| CN111523976A (en) | Commodity recommendation method and device, electronic equipment and storage medium | |
| Panniello et al. | Incorporating context into recommender systems: an empirical comparison of context-based approaches | |
| US9082086B2 (en) | Adaptively learning a similarity model | |
| CN105469263A (en) | Commodity recommendation method and device | |
| CN106251174A (en) | Information recommendation method and device | |
| CN111815415A (en) | Commodity recommendation method, system and equipment | |
| CN110008397B (en) | Recommendation model training method and device | |
| CN110930223A (en) | Recommendation recall method, device and storage medium based on field-aware factorization machine | |
| US10380123B2 (en) | Store-specific recommendation engine | |
| US20120296776A1 (en) | Adaptive interactive search | |
| Chen et al. | Common pitfalls in training and evaluating recommender systems | |
| CN113837842A (en) | Commodity recommendation method and equipment based on user behavior data | |
| CN113268656A (en) | User recommendation method and device, electronic equipment and computer storage medium | |
| CN114896517A (en) | Commodity recommendation method, system, equipment and storage medium | |
| Nurchayati et al. | Antecedents of user attitude towards e-commerce and future purchase intention | |
| Cao et al. | Distributed design and implementation of SVD++ algorithm for e-commerce personalized recommender system | |
| CN111654714A (en) | Information processing method, device, electronic equipment and storage medium | |
| CN111402003A (en) | System and method for realizing user-related recommendation | |
| CN108304407B (en) | Method and system for sequencing objects | |
| CN115239421A (en) | Commodity recommendation method, commodity recommendation device, commodity recommendation equipment and commodity recommendation medium | |
| Bajpai et al. | An improved dynamic e-commerce recommendation system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20200327 |
|
| WW01 | Invention patent application withdrawn after publication |