CN108984711A - A kind of personalized APP recommended method based on layering insertion - Google Patents
A kind of personalized APP recommended method based on layering insertion Download PDFInfo
- Publication number
- CN108984711A CN108984711A CN201810742778.8A CN201810742778A CN108984711A CN 108984711 A CN108984711 A CN 108984711A CN 201810742778 A CN201810742778 A CN 201810742778A CN 108984711 A CN108984711 A CN 108984711A
- Authority
- CN
- China
- Prior art keywords
- app
- user
- level
- matching
- 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.)
- Granted
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a kind of personalized APP recommended methods based on layering insertion, mainly using strategies such as the layering of user's fine granularity, the layering of APP fine granularity, the matchings of user's-APP interlayer, the efficiency and accuracy of APP recommendation are promoted, recommends to be best suitable for the APP of its individual demand to user.Fine granularity Stratified Strategy can reduce the range of groups of users size and APP interested, therefore being capable of the more efficient demand for accurately predicting user and progress personalized recommendation;In addition, hierarchical information is relatively stable, the APP updated suitable for data dynamic recommends scene.This patent achievement can provide good user experience for smart phone user;Efficiency of service and quality are promoted for APP application market;The APP that oneself is more rapid and better promoted for APP developer is provided conveniently.
Description
Technical field
The present invention relates to a kind of personalized APP recommended methods based on layering insertion, belong to software technology field.
Background content
The application market of existing APP service providing platform such as Android and the App Store of apple, can search according to user
Rope keyword returns to related APP, but these are all based on the recommendation that traditional collaborative filtering method carries out, without being to provide
User-customized recommended.Existing APP recommends main record and journal file depending on the user's operation to be recommended, and is recommended
Be using the APP that user is relatively more or popularization degree is relatively high.Classification and the user of individual consumer and APP application are not considered
The individualized fit degree of the keyword of input, most of recommendation results are not accurate enough, recommend efficiency lower.And on the other hand,
Complicated Generalization bounds are difficult to promote in the limited mobile device of computing capability.Therefore, there is an urgent need to a kind of intelligence of lightweight
It can cell phone application Generalization bounds.
Explanation of nouns: the matched feature of level: the i.e. specific matching relationship of user and app.
Summary of the invention
The present invention overcomes the shortcomings of the prior art, and the invention discloses a kind of personalized APP based on layering insertion
Recommended method.The present invention is using plans such as the layering of user's fine granularity, the layering of APP fine granularity, the matching of user's-APP level, layering insertions
Slightly, recommend to be best suitable for its individual demand to user by improving recommendation quality to promote the efficiency and accuracy of APP recommendation
APP.
In order to solve the above technical problems, the technical scheme adopted by the invention is as follows:
A kind of personalized APP recommended method based on layering insertion, which comprises the steps of:
Step 1: user data and APP data are obtained;
Step 2: carrying out distinguishing hierarchy: determining user and APP respectively according to the attributive character of user data and APP data
Respective hierarchical relationship;Successively refinement divides level, and determines incidence relation respectively between layers;
Step 3: carrying out the matching of user-APP feature level and preferentially chooses: according to the attributive character of user and APP and
The resulting hierarchical relationship of step 2 carries out the matching of user-APP level, and obtains each use from user data and APP data
- APP level matched initial matching feature in family establishes user's-APP score in predicting model, using machine learning model to level
Matched initial characteristics are trained, according to model prediction as a result, the selection matched relevance of level is greater than given threshold
Level matching characteristic is added to feature list;
Score_List=ML_predict (X1,X2…XN)
What Score_List was represented is the marking list to the matched feature of user's-APP level, what ML_predict was represented
It is the machine learning prediction model used in an experiment, what X was represented is the feature in level matching, and what N was represented is level matching
The number of middle feature;
Q={ M1,M2…Mn}
That q is represented is the spy for obtaining marking by machine learning score in predicting model and being higher than in the level matching of given threshold
It is feature list that collection, which is closed,;M represents the feature that marking is higher than in the level matching of given threshold, and n represents the number of feature;
Step 4: it is embedded in the realization of inquiry: including the following steps:
One), the searching keyword of user is matched with the matching characteristic that step 3 level matches to obtain in feature list,
Multiple user-APP the levels for being higher than threshold value with user query Keywords matching degree are selected, and match acquisition with level in step 3
The marking of level combine to obtain screening value, the feature and APP number in corresponding level are selected, so that it is higher to select screening value
Preceding a levels in APP as level match key feature;
Two) insertion inquiry, is implemented as follows:
Step 3 obtain characteristic set carried out as dictionary using;Wherein determine corresponding A PP and use in level matching
Weight between each layer in family, the maximum characteristic storage of weight between corresponding A PP and each layer of user in the matching of dictionary storage tier, to every
One layer of keyword is handled, and what is stored in dictionary is the key feature in level matching;The quantity and dictionary of key feature
Size constantly extending and increasing using the time with user's-APP score in predicting model, be dynamically generated in one
Journey, specifically, calculating the collective formula (1) of key feature present in dictionary:
Φ () function calculates the set of existing level matching key feature in output dictionary, and q represents the spy in step 3
List is levied, w represents the keyword of inquiry;| q | expression be key feature in dictionary number;Query statement is embedded in, is passed through
Key feature matching degree in keyword and dictionary in comparison query sentence recommends the keyword with input inquiry to user
The preceding N APP application of matching degree from high to low.
It is further to improve, N < 10.
It is further to improve, in the step 2, the attributive character of user data include the occupation of user, the age, region,
Gender;The attributive character of APP information includes APP number, classification, comment, download, click volume.
It is further to improve, the method for carrying out distinguishing hierarchy are as follows: most by broad scope in the data of user and the data of APP
The secondary wide word of the broad scope that wide word includes to the word of the first level as the first level, then as the second level, according to
It is secondary to analogize;Such as occupation and age are the first level in the data of user;Occupation includes student, teacher, and the age includes growing up, not
Adult;Then student, teacher and adult, it is teenage be the second level;Student includes university student, middle school student and high school student;Adult
Including young and old, then university student, middle school student, high school student, youth and old age are third level.
It is further to improve, by keyword weight computational algorithm determine in level matching corresponding A PP and each layer of user it
Between weight.
It is further to improve, in the step 4, select the multiple use for being higher than threshold value with user query Keywords matching degree
When-APP the level of family, the threshold value is set as the average value of all user-APP level matching marking.
Further to improve, in the step 4, user-APP level matches the level obtained with level in step 3
The mode that marking combines are as follows: user query keyword and the matched matching degree of user-APP level add user-APP level
Screening value is obtained with value;A < 20.
Detailed description of the invention
Fig. 1 is the APP suggested design flow chart based on Stratified Strategy;
Fig. 2 is the hierarchical diagram of user;
Fig. 3 is the hierarchical diagram of APP;
Fig. 4 is layering insertion prognostic chart;
Fig. 5 is level matching figure.
Specific embodiment
This patent recommends overall plan as follows based on the personalized APP of layering insertion:
Firstly the need of the data for obtaining user and APP, the data of user mainly include the occupation of user, the age, region, disappear
Flat, hobby of water wasting etc..The information of APP mainly includes the information such as APP classification, scoring, comment, download, click volume.Because original
Data have redundancy, so pre-processing firstly the need of to data, filter out extraneous data.Then, to user and APP into
Row layered shaping successively refines user and APP information, and determines relationship between layers.Then, user and APP are obtained
Between level related information, hierarchical relationship is analyzed, the range of user and the classification of APP delimited to fine granularity, for
It is contacted between user and APP, then by the query statement and keyword of layering insertion user, the acquisition of keyword is mainly logical
It crosses natural language processing and machine learning method carries out data processing, extract keyword, layering incorporation model, which is one, to be expanded
The model of exhibition, by the relationship between user's-APP- query statement, to achieve the purpose that personalized recommendation APP.
Fig. 1 illustrates the overall plan of this patent.Fig. 2 and Fig. 3 respectively shows this patent to the fine granularity of user and APP
Layered shaping.Fig. 4 illustrates layering insertion prediction flow chart.
One, data processing
It realizes based on the personalized APP recommended models of layering insertion, needs to clean data and handled.We are first
User and APP application data are subjected to hierarchical classification.
1. user stratification
Demand and hobby of the different class of subscribers to APP are all not quite similar.For example, the user preferences of different age group have
Institute is different, and the user demand of different occupation is also different.Equally, the information such as gender, distributional region can also cause the selection of user
It influences.This patent carries out user stratification substantially according to 4 kinds of information such as occupation, age, gender, region, as shown in table 1.
1. user stratification of table
2.APP layering
Current APP substantial amounts, it is many kinds of.Existing various APP application shops adopt the application of such multi-quantity
Classification is taken.This patent refers to existing classification information, and carries out micronization processes.There are many type of APP, in one species again
It has a various a large amount of APP, how to recommend the APP for being best suitable for its demand and hobby not only urgent but also very challenging to user.
APP has been done the layering of refinement by this patent, as shown in table 2.
Table 2.APP layering
3. insertion inquiry
Mainly handled using machine learning method.By handling the relationship between user and APP, mainly
The relationship between user and application is determined to grading of the download of APP, the keyword of comment, APP etc. by user.
Dictionary q storage is the maximum characteristic storage of weight between corresponding A PP and each layer of user in level matching, to each
The keyword of layer is handled, and what is stored in dictionary is the key feature in level matching;The quantity of key feature and dictionary
Size constantly can extend and increase, and be dynamically generated process in one.Dictionary is obtained using following formula (2)
Present in key feature set:
φ function calculates the set of key feature in existing level matching in output dictionary, and q represents the spy in step 3
List is levied, w represents the keyword of inquiry.| q | expression be key feature in dictionary number.Query statement is embedded in, is passed through
Key feature matching degree in keyword and dictionary in comparison query sentence recommends the keyword with input inquiry to user
The preceding N APP of matching degree from high to low.
Wherein: level matching characteristic refers to the feature obtained in relatively high level of giving a mark in step 3.Two, level is associated with
1. user grouping
User is grouped by this patent by hierarchical information similarity first.Such as: between 18 years old to 26 years old at the age
University student, and school location is grouped recommendation as a group in the same city.It is layered by fine granularity similar
Degree can find the groups of users for being more likely to same or similar demand, carry out a group recommendation to such group, can increase and push away
The efficiency and accuracy recommended, promote the quality of recommendation.Grouping citing is shown in Table 3.
The citing of 3. user grouping of table
Group 1 | University student, 18-25, male, Guangdong ... |
Group 2 | University teacher, 25-40, female, Hunan ... |
Group 3 | White collar, 20-28, female, Shanghai ... |
It is other | … |
User is grouped by fine granularity layering, on the one hand can reduce the size of groups of users, it on the other hand can also
To reduce the range of APP interested to groups of users.On this basis, in conjunction with the historical data of user, user behavior, user
The information such as click volume, download time, number of reviews to APP carry out personalization using the collaborative filtering based on user and push away
It recommends, so as to obtain the corresponding most interested APP list of the group.
2. level matches
Method), main thought is: if some word or phrase by carrying out fine granularity layering to user and APP, can
Recommended range is reduced, improves and recommends efficiency and accuracy.In addition, while layering, it was noted that being deposited between different levels
There are useful connections between certain association, especially client layer and APP layers.The class weight of different layers is not yet
Together.The calculating of weight use similar TF-IDF (it is high (i.e. TF high) that keyword weight calculates the frequency occurred in an article, and
The number occurred in other articles is few (IDF high), then thinks that this word or phrase have good class discrimination ability.Than
Such as, the APP for aspect of doing shopping recommends, and gender is that woman's weighing factor possibility will be bigger in client layer;At the same time, to specific
Such as 25-40 years old working women of crowd (in general like travel and do shopping), travel in APP and the weight of shopping will be compared with
It is high.Therefore, when carrying out APP recommendation, a preliminary screening can be carried out according to the matching between level.After grouping,
The higher class hierarchy of weight ratio is preferentially selected in respective layer.Possibility connection between layers is as shown in Figure 5.
In user and APP layers of corresponding relationship, client layer each single item has corresponded to weight possessed by APP each single item
Difference, therefore the side connected has different weights.When obtaining user information, matching point can be first carried out with the higher APP of weight
Analysis, each layer is all preferentially found and user information matches the highest feature of weight.Equally, APP information is also similarly operated,
It finds out the highest feature of weight and carries out priority match.By finding the highest feature of weight, most possibly like to obtain user
APP, with this achieve the effect that reduce recommended range, promoted recommend accuracy.Weighting can be used in relationship between layers
Figure indicates.Level weight example is as shown in table 4 (assuming that proportion range is 1-5).
4. level weight example of table
Music | Shopping | Study | |
Student | 5 | 4 | 5 |
Adult | 3 | 4 | 3 |
Female | 4 | 5 | 4 |
Guangdong | 3 | 5 | 4 |
Four, personalized recommendation and model analysis
Front is then to apply relationship embedding user-the fine granularity layering of user and APP information and level association process
Enter query statement, realize that personalized accurate APP recommends, realizes the combination of user-APP- inquiry three, utilize machine learning
The scalability of method realization personalized recommendation.Based on above-mentioned processing result, we can combine existing collaborative filtering
Technology, machine learning techniques, design the personalized APP recommended models of efficient lightweight grade, and test recommendation results and effect
Card analysis.
Examples detailed above is only the specific embodiment of the present invention, is also being sent out its simple transformation, replacement etc.
In bright protection scope.
Claims (7)
1. a kind of personalized APP recommended method based on layering insertion, which comprises the steps of:
Step 1: user data and APP data are obtained;
Step 2: carrying out distinguishing hierarchy: determining user and APP respectively respectively according to the attributive character of user data and APP data
Hierarchical relationship;Successively refinement divides level, and determines incidence relation respectively between layers;
Step 3: carrying out the matching of user-APP feature level and preferentially chooses: according to the attributive character and step of user and APP
Two resulting hierarchical relationships carry out the matching of user-APP level, and obtain each user-from user data and APP data
The matched initial matching feature of APP level establishes user's-APP score in predicting model, is matched using machine learning model to level
Initial characteristics be trained, according to model prediction as a result, selection the matched relevance of level be greater than given threshold level
Matching characteristic is added to feature list;
Score_List=ML_predict (X1,X2…XN)
What Score_List was represented is marking list to the matched feature of user's-APP level, what ML_predict was represented be
The machine learning prediction model used in experiment, what X was represented is the feature in level matching, and what N was represented is special in level matching
The number of sign;
Q={ M1,M2…Mn}
What q was represented is the feature set for obtaining marking by machine learning score in predicting model and being higher than in the level matching of given threshold
Closing is feature list;M represents the feature that marking is higher than in the level matching of given threshold, and n represents the number of feature;
Step 4: it is embedded in the realization of inquiry: including the following steps:
One), the searching keyword of user is matched with the matching characteristic that step 3 level matches to obtain in feature list, is selected
It is higher than multiple user-APP levels of threshold value with user query Keywords matching degree, and matches the layer obtained with level in step 3
The marking of grade combines to obtain screening value, selects feature in corresponding level and APP number, thus select screening value it is higher before
APP in a levels matches key feature as level;
Two) insertion inquiry, is implemented as follows:
Step 3 obtain characteristic set carried out as dictionary using;Wherein determine that corresponding A PP and user are each in level matching
Weight between layer, the maximum characteristic storage of weight between corresponding A PP and each layer of user in the matching of dictionary storage tier, to each layer
Keyword handled, stored in dictionary be level matching in key feature;The quantity of key feature and dictionary it is big
Small constantly being extended and being increased using the time with user's-APP score in predicting model, is dynamically generated process in one, has
Body, calculate the collective formula (1) of key feature present in dictionary:
Φ () function calculates the set of existing level matching key feature in output dictionary, and q represents the characteristic series in step 3
Table, w represent the keyword of inquiry;| q | expression be key feature in dictionary number;Query statement is embedded in, by comparing
Key feature matching degree in keyword and dictionary in query statement recommends the Keywords matching with input inquiry to user
The preceding N APP application of degree from high to low.
2. the personalized APP recommended method as described in claim 1 based on layering insertion, which is characterized in that N < 10.
3. the personalized APP recommended method as described in claim 1 based on layering insertion, which is characterized in that the step 2
In, the attributive character of user data includes the occupation of user, age, region, gender;The attributive character of APP information includes that APP is compiled
Number, classification, comment, download, click volume.
4. the personalized APP recommended method as claimed in claim 3 based on layering insertion, which is characterized in that carry out level and draw
The method divided are as follows: using the most wide word of broad scope in the data of user and the data of APP as the first level, then to first layer
The secondary wide word of the broad scope that the word of grade includes as the second level, and so on;Such as occupation and year in the data of user
Age is the first level;Occupation includes student, teacher, and the age includes growing up, being teenage;Then student, teacher and adult, it is teenage
For the second level;Student includes university student, middle school student and high school student;Adult include it is young and old, then university student, middle school student,
High school student, youth and old age are third level.
5. the personalized APP recommended method as described in claim 1 based on layering insertion, which is characterized in that pass through keyword
Weight calculation algorithm determines weight between corresponding A PP and each layer of user in level matching.
6. the personalized APP recommended method as described in claim 1 based on layering insertion, which is characterized in that the step 4
In, select with user query Keywords matching degree be higher than threshold value multiple user-APP levels when, the threshold value is set as all
The average value of user's-APP level matching marking.
7. the personalized APP recommended method as described in claim 1 based on layering insertion, which is characterized in that the step 4
In, user-APP level is the same as the mode that combines of marking for the level that level matching obtains in step 3 are as follows: user query key
Word and the matched matching degree of user-APP level are plus user-APP level matching value acquisition screening value;A < 20.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810742778.8A CN108984711B (en) | 2018-07-09 | 2018-07-09 | Personalized APP recommendation method based on hierarchical embedding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810742778.8A CN108984711B (en) | 2018-07-09 | 2018-07-09 | Personalized APP recommendation method based on hierarchical embedding |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108984711A true CN108984711A (en) | 2018-12-11 |
CN108984711B CN108984711B (en) | 2020-09-11 |
Family
ID=64537606
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810742778.8A Active CN108984711B (en) | 2018-07-09 | 2018-07-09 | Personalized APP recommendation method based on hierarchical embedding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108984711B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114493377A (en) * | 2022-04-06 | 2022-05-13 | 广州平云小匠科技有限公司 | Work order dispatching method and system |
CN115146741A (en) * | 2022-08-02 | 2022-10-04 | 统信软件技术有限公司 | Operating system level feature determination method, computing device and storage medium |
CN117435817A (en) * | 2023-12-20 | 2024-01-23 | 泰安北航科技园信息科技有限公司 | BI intelligent center system based on industry big data |
CN117454018A (en) * | 2023-12-22 | 2024-01-26 | 深圳市康莱米电子股份有限公司 | Education platform implementation method and system based on tablet personal computer |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103049468A (en) * | 2011-10-11 | 2013-04-17 | 谷歌公司 | Determining intent of a recommendation on a mobile application |
US20170270432A1 (en) * | 2016-03-17 | 2017-09-21 | Accenture Global Solutions Limited | System modernization using machine learning |
CN107818492A (en) * | 2017-10-10 | 2018-03-20 | 平安科技(深圳)有限公司 | Products Show device, method and computer-readable recording medium |
-
2018
- 2018-07-09 CN CN201810742778.8A patent/CN108984711B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103049468A (en) * | 2011-10-11 | 2013-04-17 | 谷歌公司 | Determining intent of a recommendation on a mobile application |
US20170270432A1 (en) * | 2016-03-17 | 2017-09-21 | Accenture Global Solutions Limited | System modernization using machine learning |
CN107818492A (en) * | 2017-10-10 | 2018-03-20 | 平安科技(深圳)有限公司 | Products Show device, method and computer-readable recording medium |
Non-Patent Citations (1)
Title |
---|
赵海燕等: "基于APP使用历史主题分析的个性化推荐算法", 《电子科技》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114493377A (en) * | 2022-04-06 | 2022-05-13 | 广州平云小匠科技有限公司 | Work order dispatching method and system |
CN115146741A (en) * | 2022-08-02 | 2022-10-04 | 统信软件技术有限公司 | Operating system level feature determination method, computing device and storage medium |
CN115146741B (en) * | 2022-08-02 | 2024-06-07 | 统信软件技术有限公司 | Feature determination method of operating system level, computing device and storage medium |
CN117435817A (en) * | 2023-12-20 | 2024-01-23 | 泰安北航科技园信息科技有限公司 | BI intelligent center system based on industry big data |
CN117435817B (en) * | 2023-12-20 | 2024-03-15 | 泰安北航科技园信息科技有限公司 | BI intelligent center system based on industry big data |
CN117454018A (en) * | 2023-12-22 | 2024-01-26 | 深圳市康莱米电子股份有限公司 | Education platform implementation method and system based on tablet personal computer |
CN117454018B (en) * | 2023-12-22 | 2024-03-08 | 深圳市康莱米电子股份有限公司 | Education platform implementation method and system based on tablet personal computer |
Also Published As
Publication number | Publication date |
---|---|
CN108984711B (en) | 2020-09-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111680173A (en) | CMR model for uniformly retrieving cross-media information | |
CN103617230B (en) | Method and system for advertisement recommendation based microblog | |
CN102289522B (en) | Method of intelligently classifying texts | |
CN108984711A (en) | A kind of personalized APP recommended method based on layering insertion | |
CN108763213A (en) | Theme feature text key word extracting method | |
CN110532379B (en) | Electronic information recommendation method based on LSTM (least Square TM) user comment sentiment analysis | |
CN109960786A (en) | Chinese Measurement of word similarity based on convergence strategy | |
CN106649275A (en) | Relation extraction method based on part-of-speech information and convolutional neural network | |
CN109635083B (en) | Document retrieval method for searching topic type query in TED (tele) lecture | |
CN110110225B (en) | Online education recommendation model based on user behavior data analysis and construction method | |
CN104133817A (en) | Online community interaction method and device and online community platform | |
CN110209808A (en) | A kind of event generation method and relevant apparatus based on text information | |
CN103778227A (en) | Method for screening useful images from retrieved images | |
CN105426514A (en) | Personalized mobile APP recommendation method | |
CN113076476B (en) | User portrait construction method of microblog heterogeneous information | |
CN104778283B (en) | A kind of user's occupational classification method and system based on microblogging | |
WO2018176913A1 (en) | Search method and apparatus, and non-temporary computer-readable storage medium | |
CN107784110A (en) | A kind of index establishing method and device | |
CN103778206A (en) | Method for providing network service resources | |
CN106886512A (en) | Article sorting technique and device | |
CN105512333A (en) | Product comment theme searching method based on emotional tendency | |
CN108647800A (en) | A kind of online social network user missing attribute forecast method based on node insertion | |
CN103761286B (en) | A kind of Service Source search method based on user interest | |
CN106126605A (en) | A kind of short text classification method based on user's portrait | |
CN108920521A (en) | User's portrait-item recommendation system and method based on pseudo- ontology |
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 | ||
CB03 | Change of inventor or designer information | ||
CB03 | Change of inventor or designer information |
Inventor after: Jiang Wenjun Inventor after: Liu Dong Inventor after: Ren Desheng Inventor before: Jiang Wenjun Inventor before: Liu Dong |
|
GR01 | Patent grant | ||
GR01 | Patent grant |