CN108984711B

CN108984711B - Personalized APP recommendation method based on hierarchical embedding

Info

Publication number: CN108984711B
Application number: CN201810742778.8A
Authority: CN
Inventors: 姜文君; 刘栋; 任德盛
Original assignee: Hunan University
Current assignee: Hunan University
Priority date: 2018-07-09
Filing date: 2018-07-09
Publication date: 2020-09-11
Anticipated expiration: 2038-07-09
Also published as: CN108984711A

Abstract

The invention discloses a personalized APP recommendation method based on hierarchical embedding, which mainly adopts strategies such as user fine-grained layering, APP fine-grained layering, user-APP interlayer matching and the like to improve the efficiency and accuracy of APP recommendation and recommend an APP which best meets the personalized requirements of a user. The fine-grained hierarchical strategy can reduce the size of a user group and the range of interested APP, so that the requirements of the user can be predicted more efficiently and accurately and personalized recommendation can be carried out; in addition, the hierarchical information is relatively stable, and the method is suitable for APP recommendation scenes of data dynamic updating. The patent result can provide good user experience for the smart phone user; service efficiency and quality are improved for the APP application market; the method provides convenience for APP developers to popularize own APP more quickly and better.

Description

Personalized APP recommendation method based on hierarchical embedding

Technical Field

The invention relates to a personalized APP recommendation method based on hierarchical embedding, and belongs to the technical field of software.

Background content

Existing APP service providing platforms such as an application market of android and an APP Store of apple can return related APPs according to user search keywords, but these are recommendations based on a conventional collaborative filtering method, rather than providing user personalized recommendations. The conventional APP recommendation is mainly based on operation records and log files of users, and the recommended APP is more users or high in popularization degree. The categories of the individual users and the APP and the personalized matching degree of the keywords input by the users are not considered, most recommendation results are not accurate enough, and the recommendation efficiency is low. On the other hand, complex recommendation strategies are difficult to popularize on mobile devices with limited computing power. Therefore, a lightweight smartphone APP recommendation strategy is urgently needed.

The noun explains: characteristics of hierarchical matching: i.e. the specific matching relationship of the user and the app.

Disclosure of Invention

The invention overcomes the defects in the prior art and discloses a personalized APP recommendation method based on hierarchical embedding. The method adopts strategies such as user fine-grained layering, APP fine-grained layering, user-APP level matching, layered embedding and the like to improve the efficiency and accuracy of APP recommendation, and recommends the APP which best meets the personalized requirements of the user by improving the recommendation quality.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows:

a personalized APP recommendation method based on hierarchical embedding is characterized by comprising the following steps:

the method comprises the following steps: obtaining user data and APP data;

step two, hierarchical division is carried out: respectively determining respective hierarchical relations of the user and the APP according to the attribute characteristics of the user data and the APP data; thinning layer by layer, dividing the hierarchy, and determining the association relationship between each layer;

step three: carrying out user-APP feature level matching and preferred selection: according to the attribute characteristics of the user and the APP and the hierarchical relationship obtained in the second step, carrying out user-APP hierarchical matching, obtaining initial matching characteristics of each user-APP hierarchical matching from the user data and the APP data, establishing a user-APP score prediction model, training the initial characteristics of the hierarchical matching by using a machine learning model, and adding the hierarchical matching characteristics of which the relevance is greater than a set threshold value into a characteristic list according to the result of model prediction;

Score_List＝ML_predict(X₁,X₂…X_N)

score _ List represents a scoring List of features for user-APP hierarchical matching, ML _ predict represents a machine learning prediction model adopted in an experiment, X represents features in hierarchical matching, and N represents the number of features in hierarchical matching;

q＝{M₁,M₂…M_n}

q represents a feature list which is a feature set in hierarchical matching with a score higher than a set threshold value and is obtained through a machine learning score prediction model; m represents the features in the hierarchical matching with the score higher than the set threshold value, and n represents the number of the features;

step four: implementation of the embedded query: the method comprises the following steps:

firstly), matching the query key words of the users with the matching features in the feature list obtained by the three-level matching in the step, selecting a plurality of user-APP levels with the matching degree higher than a threshold value with the query key words of the users, combining the levels with the grades obtained by the matching of the middle level in the step to obtain a screening value, and selecting the features and APP numbers in the corresponding levels so as to select the APP in the front a-name level with higher screening value as the key feature for level matching;

second) embedding query, specifically implementing as follows:

using the feature set obtained in the third step as a word stock; determining weights between corresponding APPs and user layers in hierarchical matching, storing features with the largest weights between the corresponding APPs and the user layers in thesaurus storage hierarchical matching, and processing keywords of each layer, wherein the key features in the hierarchical matching are stored in the thesaurus; the number of key features and the size of the lexicon are continuously expanded and increased along with the use time of the user-APP score prediction model, and the method is in a dynamic generation process, and specifically, a set formula (1) of the key features existing in the lexicon is calculated:

calculating a set of existing hierarchical matching key features in an output word library by using a phi () function, wherein q represents a feature list in the third step, and w represents a query key word; the | q | represents the number of key features in the lexicon; and embedding the query sentence, and recommending the APP with the top N names from high to low matching degree with the keywords of the input query to the user by comparing the matching degree of the keywords in the query sentence with the key features in the word stock.

In a further improvement, N < 10.

In a further improvement, in the second step, the attribute characteristics of the user data include occupation, age, region and gender of the user; the attribute characteristics of the APP information comprise APP numbers, categories, comments, download amount and click amount.

Further improvement, the method for hierarchical division comprises the following steps: taking the word with the widest generalization range in the data of the user and the data of the APP as a first level, then taking the words with the widest generalization range contained in the words of the first level as a second level, and so on; for example, the occupation and age in the user's data is at a first level; profession includes students, teachers, age includes adult, minor; then the student, teacher and adult, minor are all level two; students include college students, middle school students and high school students; adults include young and old, and college students, middle school students, high school students, young and old are all at the third level.

And further improvement, determining the weight between each layer of the corresponding APP and the user in the hierarchical matching through a keyword weight calculation algorithm.

In the fourth step, when a plurality of user-APP levels with the matching degree with the user query keyword higher than a threshold value are selected, the threshold value is set as an average value of matching scores of all the user-APP levels.

In a further improvement, in the fourth step, the user-APP level is combined with the score of the level obtained by level matching in the third step in a manner that: the matching degree of the user query keyword and the user-APP level matching is added with the user-APP level matching value to obtain a screening value; a is less than 20.

Drawings

FIG. 1 is a flowchart of an APP recommendation scheme based on a hierarchical policy;

FIG. 2 is a user's hierarchy;

FIG. 3 is a layered structure diagram of APP;

FIG. 4 is a diagram of hierarchical embedding prediction;

fig. 5 is a hierarchical matching diagram.

Detailed Description

The patent recommends the overall scheme as follows based on layering embedding's individualized APP:

firstly, data of a user and APP (application) are required to be obtained, and the data of the user mainly comprises occupation, age, region, consumption level, hobbies and the like of the user. The information of the APP mainly comprises information such as APP categories, scores, comments, download amount and click amount. Because the original data is redundant, the data needs to be preprocessed first to filter out extraneous data. And then, carrying out layered processing on the user and the APP, refining the information of the user and the APP layer by layer, and determining the relation between layers. Then, obtaining hierarchical association information between the user and the APP, analyzing the hierarchical relationship, finely dividing the range of the user and the type of the APP, aiming at the relation between the user and the APP, then embedding query sentences and key words of the user in a hierarchical manner, obtaining the key words mainly comprises data processing through natural language processing and machine learning methods, extracting key words, wherein the hierarchical embedding model is an extensible model, and achieving the purpose of individually recommending the APP through the relationship between the user and the APP-query sentences.

Figure 1 shows the overall scheme of the patent. Fig. 2 and fig. 3 show fine-grained layered processing of the user and APP by the present patent, respectively. Fig. 4 illustrates a hierarchical embedded prediction flow diagram.

Data processing

The realization of the personalized APP recommendation model based on the hierarchical embedding requires the cleaning and processing of data. We first classify user and APP application data hierarchically.

1. User layering

The demands and preferences of different user classes for APP vary. For example, the preferences of users in different age groups are different, and the needs of users in different professions are also different. Similarly, information such as gender and distribution area also affects the user's selection. The patent roughly classifies users according to 4 kinds of information such as occupation, age, sex, region, and the like, as shown in table 1.

TABLE 1 user stratification

Layering of APP

The number of APP is huge at present, and the variety is various. Various APP stores have adopted such a large number of applications. The patent refers to the existing classification information and carries out refinement processing. The types of the APP are various, and a large number of APPs exist in the same type, so that how to recommend the APP which best meets the needs and preferences of users is urgent and very challenging. This patent makes a refined layering of APPs, as shown in table 2.

TABLE 2 layering of APP

3. Embedded queries

The processing is mainly performed using a machine learning method. The relation between the user and the APP is processed, and the relation between the user and the application is mainly determined through the downloading amount of the APP by the user, keywords of comments, the rating of the APP and the like.

The word bank q stores feature storage with the maximum weight between corresponding APP and each layer of the user in hierarchical matching, keywords of each layer are processed, and key features in the hierarchical matching are stored in the word bank; the number of key features and the size of the lexicon are continuously expanded and increased, and the generation process is dynamic. The set of key features present in the lexicon is obtained using the following formula (2):

phi function calculates the set of key features in the existing hierarchical matching in the output word library, q represents the feature list in step three, and w represents the keywords of the query. | q | represents the number of key features in the lexicon. And embedding the query sentence, and recommending the top N APPs with the matching degree from high to low with the keyword of the input query to the user by comparing the matching degree of the keyword in the query sentence with the key characteristic in the word stock.

Wherein: the level matching features refer to the features in the level with higher score obtained in step three. Two, hierarchical association

1. User grouping

The patent firstly groups users according to the similarity of layered information. For example: grouping recommendations are made for college students between 18 and 26 years of age, with school locations in the same city, as a group. User groups with the same or similar requirements can be found through the fine-grained hierarchical similarity, group recommendation is carried out on the groups, the recommendation efficiency and accuracy can be improved, and the recommendation quality is improved. See table 3 for an example of the groupings.

TABLE 3 user grouping example

Group 1	College student, 18-25, male, Guangdong, …
		Group 2	University teacher, 25-40, woman, Hunan, …
Group 3	White-collar, 20-28 parts, female, Shanghai, … parts
		Others	…

The users are grouped according to fine-grained hierarchy, so that the size of the user group can be reduced on one hand, and the range of the APP interested by the user group can be reduced on the other hand. On the basis, information such as historical data of the user, user behaviors, clicking amount of the APP by the user, downloading times, comment quantity and the like is combined, personalized recommendation is performed by using a collaborative filtering algorithm based on the user, and therefore the APP list which is most interested in and corresponds to the group can be obtained.

2. Hierarchical matching

Method), the main idea is: if a certain word or phrase is layered with the user and the APP in a fine-grained manner, the recommendation range can be narrowed, and the recommendation efficiency and accuracy can be improved. In addition, while layering, we note that there is some correlation between different layers, especially there is a useful connection between the user layer and the APP layer. The class weights of different layers also differ. For example, in the aspect of APP recommendation in shopping, the influence weight of a female with gender in a user layer may be larger, meanwhile, the weight of tourism and shopping in APP recommendation in a specific group such as professional females (generally like tourism and shopping), and the weight of tourism and shopping in APP is higher, therefore, when the APP recommendation is carried out, primary screening can be carried out according to matching between levels, after grouping, the category level with higher weight in the corresponding layer is preferentially selected, and possible connection between layers is shown in figure 5.

In the corresponding relation between the user and the APP layer, each item of the user layer has different weights corresponding to each item of the APP, so that the connected edges have different weights. When user information is obtained, matching analysis is performed with APP with higher weight, and each layer preferentially searches for the features with the highest matching weight with the user information. Similarly, the APP information also performs the same operation, and finds out the features with the highest weight for preferential matching. The APP most likely liked by the user is obtained by searching the features with the highest weight, so that the effects of reducing the recommendation range and improving the recommendation accuracy are achieved. The relationships between layers may be represented using weighted graphs. The hierarchical weight representation is shown in Table 4 (assuming a weight range of 1-5).

TABLE 4 example hierarchy weights

	Music	Shopping	Study of
				Student's desk	5	4	5
Adults	3	4	3
				Woman	4	5	4
Guangdong (Chinese character of Guangdong)	3	5	4

Fourth, personalized recommendation and model analysis

The method comprises the steps of performing fine-grained layering and hierarchical association processing on user and APP information, then embedding a user-application relation into an inquiry statement, realizing personalized accurate APP recommendation, realizing the combination of the user-APP-inquiry, and realizing the extensibility of the personalized recommendation by using a machine learning method. Based on the processing results, an efficient and lightweight personalized APP recommendation model can be designed by combining the existing collaborative filtering technology and the machine learning technology, and the recommendation results and effects are verified and analyzed.

The above example is only one embodiment of the present invention, and simple changes, substitutions, and the like are also within the scope of the present invention.

Claims

1. A personalized APP recommendation method based on hierarchical embedding is characterized by comprising the following steps:

the method comprises the following steps: obtaining user data and APP data;

Score_List＝ML_predict(X₁,X₂…X_N)

q＝{M₁,M₂…M_n}

second) embedding query, specifically implementing as follows:

2. The personalized APP recommendation method based on hierarchical embedding of claim 1, wherein N < 10.

3. The personalized APP recommendation method based on hierarchical embedding of claim 1, wherein in the second step, the attribute features of the user data include occupation, age, region, and gender of the user; the attribute characteristics of the APP information comprise APP numbers, categories, comments, download amount and click amount.

4. The personalized APP recommendation method based on hierarchical embedding of claim 3, wherein the method for hierarchical division comprises: taking the word with the widest generalization range in the data of the user and the data of the APP as a first level, then taking the words with the widest generalization range contained in the words of the first level as a second level, and so on; specifically, the occupation and age in the user data are at a first level; profession includes students, teachers, age includes adult, minor; then the student, teacher and adult, minor are all level two; students include college students, middle school students and high school students; adults include young and old, and college students, middle school students, high school students, young and old are all at the third level.

5. The method for recommending personalized APP based on hierarchical embedding according to claim 1, wherein the weight between the corresponding APP and each layer of the user in the hierarchical matching is determined by a keyword weight calculation algorithm.

6. The personalized APP recommendation method based on hierarchical embedding of claim 1, wherein in the fourth step, when a plurality of user-APP levels with matching degree with the user query keyword higher than a threshold value are selected, the threshold value is set as an average value of matching scores of all user-APP levels.

7. The personalized APP recommendation method based on hierarchical embedding of claim 1, wherein in step four, the user-APP level is combined with the score of the level obtained by level matching in step three by: the matching degree of the user query keyword and the user-APP level matching is added with the user-APP level matching value to obtain a screening value; a is less than 20.