CN106055566A - Mobile phone game recommendation method aiming at mobile advertisement users - Google Patents

Abstract

The invention discloses a mobile phone game recommendation method aiming at mobile advertisement users. When mobile phone users use a mobile phone application, the users often click advertisements released by the mobile phone application. Advertisement clicking behaviors of the mobile phone users can reflect interest preferences of the mobile phone users to a certain extent. Through entity extraction of advertisement introduction files, an advertisement entity set preferred by the mobile phone users can be established according to advertisement clicking records of the mobile phone users. Aiming at users with game records among the mobile phone advertisement users, a game entity set and a game set preferred by the users can be further extracted, and the two sets together with the advertisement entity set can jointly reflect the users' interest preferences. Aiming at the mobile advertisement users with the game records, a neighbor user set is established through cosine similarity computation of a user model, and personalized recommendation can be carried out; and aiming at the mobile advertisement users without the game records, a cosine similarity between the users and the advertisement users with the game records is computed, and a neighbor user set is established, so that personalized recommendation can be carried out.

Description

Method is recommended in mobile phone games towards mobile advertising user

Technical field

The present invention relates to the technical field that mobile phone games are recommended, refer in particular to a kind of mobile phone towards mobile advertising user and swim Play recommendation method.

Background technology

Along with developing rapidly of the Internet, the information faced in people's daily life grows with each passing day.For solving faced by people Magnanimity information at a loss as to what to do, it is recommended that system is arisen at the historic moment.Current proposed algorithm is divided into three kinds: recommendation based on collaborative filtering System, content-based recommendation system and the commending system of mixing.Commending system based on collaborative filtering mainly includes The technology such as User-Based, Item-Based and Model-Based.Amazon shopping website mainly uses Item-Based skill Art, by the historical record analysis to user, carries out personalized recommendation to user.Content-based recommendation system, mainly structure Build the characteristic vector of article, user, carry out personalized recommendation by the similarity calculating characteristic vector.More current news websites Main use content-based recommendation system.Mixed type commending system, is by commending system based on collaborative filtering with based on interior The commending system held is combined, and the advantage drawing both has wider array of subject range.

Cellphone subscriber grows with each passing day and the captivation of mobile phone games uniqueness so that swim to cellphone subscriber's personalized recommendation Play becomes trend of the times.But, mobile phone games recommend field to have a feature different from conventional recommendation: user's Game Cycle Length, cellphone subscriber's game records are few.Therefore, mobile phone games recommendation is compared with conventional recommendation, and data have bigger openness.And In mobile Internet, moving advertising generally exists, and a lot of users usually can click on when installing and use Mobile solution To the advertisement pushed, thus become mobile advertising user.Potential mobile phone games user is excavated, for them from mobile advertising user Recommend suitable mobile phone games, while promoting mobile phone games marketing so that the propelling movement of moving advertising more precision, favorably In the win-win development promoting mobile phone games industry and moving advertising industry.

Cellphone subscriber implies the interest preference of user in the behavior of different field, and has certain dependency.Phase Mobile phone games recommendation is carried out, in conjunction with user at the behavioral data of advertisement field, energy than the user data using single field of play Preferably portray the interest of user, thus promote accuracy and the multiformity that mobile phone games are recommended further.

Summary of the invention

It is an object of the invention to overcome the deficiencies in the prior art, it is provided that a kind of mobile phone games towards mobile advertising user Recommendation method, by utilizing mobile subscriber's behavioral data at advertisement field and the behavioral data in field of play, more adds Portray user interest preference wholely, and then preferably user is modeled and game recommdation.The method has good extension Property, and can apply in other recommendation fields.

For achieving the above object, technical scheme provided by the present invention is: the mobile phone games towards mobile advertising user push away Recommend method, comprise the following steps:

1) carry out participle by advertisement and game being introduced content, utilize keyword extraction techniques based on figure, obtain The significant noun in content, i.e. entity are introduced in advertisement and game；

2) for having the mobile advertising user of game records, advertising aggregator and the game set of user is built, based on extensively Accuse and the entity in content is introduced in game, build the entity sets of preferences of user；Use for the moving advertising not having game records Family, builds the advertising aggregator of user, introduces the entity in content based on advertisement, builds the entity sets of preferences of user；

3) for having the mobile advertising user of game records, log in the frequency of game according to mobile subscriber, build mobile The game preference set of advertising user；

4) for having the mobile advertising user of game records, in conjunction with entity preference and the game set of preferences of user of user Close, build the characteristic vector of user；For there is no the mobile advertising user of game records, entity preference based on user, build The characteristic vector of user；

5) for having the mobile advertising user of game records, in conjunction with game entity, advertisement entity, game sets of preferences three Individual part calculates the neighbour user of user；For there is no the mobile advertising user of game records, based on advertisement entity from having trip The mobile advertising user of play record finds neighbour user；

6) proposed algorithm based on User-Based provides the user with game recommdation list.

In step 1) in, game is introduced data, uses the conventional participle instrument with part-of-speech tagging to carry out participle, this Game is just introduced data and is resolved into the form of document-word by sample；During participle, filter out non-nominal vocabulary；Based on Graph theory falls into a trap the method for operator node importance degree to calculate the importance of each entity node；According to entity adjacent in entity sets Relation sets up figure, and in figure, the weight on limit is all 1, then uses the random walk restarted to calculate the rank value of each node, public Formula is as shown in 1-1；

$rank\left(e_i\right)=(1-\mathrm{\α})+\mathrm{\α}*\underset{e_j\∈In\left(e_i\right)}{\Σ}rank\left(e_j\right)/\left|Out\left(e_j\right)\right|---(1-1)$

Wherein, rank (e_i) presentation-entity e_iSignificance level in a document, α represents the parameter that restarts of random walk, In (e_i) refer to entity e_iNeighborhood, | Out (e_j) | refer to entity e_jNeighbours' number；

Finally, according to the rank value of each entity, extract each game and introduce the entity of the front N of data rank value；In like manner, Same method can be used to calculate each advertisement and to introduce the top n entity of data.

In step 3) in, for having the advertising user of game records, concentrate at game data, find out each user and played Game, by the user one week game login record game more than twice, put into the game sets of preferences of user；Game set of preferences Cooperation is a part for user characteristics, embodies user's preference on game selects.

In step 4) in, for there being the advertising user of game records, by the advertisement entity set of user, game entity Set and game records merge, and portray the preference profile of user from multiple dimensions, one reflection user's entirety preference of final structure Characteristic vector, dimension is user's game entity, advertisement entity and game records length sum, and form is as follows:

User characteristics={ { game entity 1, game entity 2 ... }, { advertisement entity 1, advertisement entity 2 ... }, { game Record 1, game records 2 ... } }

For not having the advertising user of game records, the advertisement entity set of user is used to build user characteristics vector, dimension Degree is the length of user advertising entity sets, and form is as follows:

User characteristics={ advertisement entity 1, advertisement entity 2 ... }.

In step 5) in, calculate the neighbor lists of two kinds of different advertising user, for there being the advertising user of game records, The Similarity Measure of user is divided into three parts: the similarity of game entity, the similarity of advertisement entity, game records similar Degree, the similarity that every part draws is with different weights w₁,w₂,w₃Carry out aggregative weighted, calculate the most similar of user Degree；Wherein, w₁,w₂,w₃It is adjusted according to experiment needs, is initialized as 1/3,1/3,1/3, i.e. three part Similarity Measure knots Really proportion is identical；Computing formula is as shown in 5-1:

Similar(u₁,u₂)=w₁*cosine(gameEntity₁,gameEntity₂)

+w₂*cosine(adEntity₁,adEntity₂)

+w₃*cosine(gameList₁,gameList₂) (5-1)

Wherein, gameEntity is game entity set, and adEntity is advertisement entity set, and gameList is that user swims Play set of records ends；Wherein cosine is cosine similarity, and computing formula is as shown in 5-2:

$\cos ine(u_i,u_j)=\frac{u_i*u_j}{\left|\right|u_i\left|\right|*\left|\right|u_j\left|\right|}=\frac{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x}_{i,k}*x_{j,k}}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{i,k}}^2}*\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{j,k}}^2}}---(5-2)$

Finally, according to the height of cosine similarity, Top-N the neighbour user of user is found out；

For there is no the advertising user of game records, more than the advertising user that calculates this user and there is game records String similarity, builds neighbour's user list of this user, shown in the following 5-3 of computing formula:

Similar(u₁,u₂)=cosine (adEntity₁,adEntity₂) (5-3)

Wherein, u₁It is the advertising user not having game records, u₂It is the advertising user having game records, final u₁Neighbour User is the advertising user having game records.

In step 6) in, for each user, according to the game records in its neighbour user, use below equation to push away Recommend:

$Grade(u_i,{\mathrm{game}}_j)=\underset{v\∈Neigh\left(u_i\right)\∩{\mathrm{game}}_j\∈gameList\left(v\right)}{\mathrm{\Σ}}Similar(u_i,v)---(6-1)$

Wherein, Neigh (u_i) it is u_iTop-N neighbour user, gameList is user's game records；

Finally, according to the value of grade from big to small, front Top-K the game of user is found out as recommendation list.

The present invention compared with prior art, has the advantage that and beneficial effect:

1, for not having the mobile advertising user of game records, the present invention utilizes these certain customers and has game records The similarity of mobile advertising user carries out mobile phone games recommendation, overcomes the cold start-up problem in conventional mobile phone game recommdation.

2, for having the advertising user of game records, the present invention is from the advertisement entity preference of user, game entity preference And game sets of preferences set out, more complete features user preference, is faced overcoming conventional mobile phone game recommdation algorithm Sparse sex chromosome mosaicism simultaneously, improve accuracy and the multiformity of recommendation.

Accompanying drawing explanation

Fig. 1 is that method flow diagram is recommended in the mobile phone games of invention.

Fig. 2 is advertisement or the game entity extraction flow chart of the present invention.

Fig. 3 is that document-entity bipartite graph is introduced in the advertisement of the present invention.

Detailed description of the invention

Below in conjunction with specific embodiment, the invention will be further described.

As it is shown in figure 1, the mobile phone games towards mobile advertising user described in the present embodiment recommend method, including following step Rapid:

1) carry out participle by advertisement and game being introduced content, utilize keyword extraction techniques based on figure, obtain The significant noun in content, i.e. entity are introduced in advertisement and game.

Entity refers to the noun that some are significant, can be that some are towards code name (such as three states, the Warring states), name in gaming (Cao behaviour, Song Jiang), place name (Chibi) etc.；Advertisement field can be some field nouns (as finance, physical culture), role (tide father, Peppery mother) etc..Game is introduced data, uses conventional participle instrument such as HandLP, OpenCLAS etc. with part-of-speech tagging to enter Row participle, thus introduces game data and resolves into the form of document-word.During participle, filter out non-nominal Vocabulary.Now the most more due to the entity number of document, therefore calculate each based on the fall into a trap method of operator node importance degree of graph theory The importance of entity node.Setting up figure according to the syntopy of entity in entity sets, in figure, the weight on limit is all 1, then makes Calculate the rank value of each node with the random walk restarted, formula is as shown in 1-1.Finally, according to the rank of each entity Value, extracts each game and introduces the entity of the front N of data rank value.Use same method to calculate each advertisement and introduce data Top n entity.The flow process of this step is as shown in Figure 2.

$rank\left(e_i\right)=(1-\mathrm{\α})+\mathrm{\α}*\underset{e_j\∈In\left(e_i\right)}{\Σ}rank\left(e_j\right)/\left|Out\left(e_j\right)\right|---(1-1)$

Wherein, rank (e_i) presentation-entity e_iSignificance level in a document, α represents the parameter that restarts of random walk, In (e_i) refer to entity e_iNeighborhood, | Out (e_j) | refer to entity e_jNeighbours' number.

2) for having the mobile advertising user of game records, advertising aggregator and the game set of user is built, based on extensively Accuse and the entity in content is introduced in game, build the entity sets of preferences of user；Use for the moving advertising not having game records Family, builds the advertising aggregator of user, introduces the entity in content based on advertisement, builds the entity sets of preferences of user.

User in moving advertising is divided into two kinds: a kind of user being to have game records；Another kind is note of not playing The user of record.For two kinds of different advertising user, the recommendation method that we use is different.But in advertisement entity preference Structure on, use identical method.

If a user is interested in certain advertisement entity, there is a strong possibility, and this advertisement entity of property appears in this use Document is introduced in the multiple advertisements clicked at family.Ad click record according to cellphone subscriber, builds advertisement-stereogram, such as Fig. 3 Shown in.Wherein summit includes advertisement vertex v_a, and entity vertex v_e, and two class summits between limit E, the weight calculation on limit is adopted Use formula 2-1.

$w(v,v^{\′})=\left\{\begin{array}{c}rank(e,a),v\∈v_a,v^{\′}\∈v_e\\ 1/out\left(v\right),v\∈v_e,v^{\′}\∈v_a\\ 0,other\end{array}\right.---(2-1)$

Limit E (the v of advertisement-inter-entity_a,v_e) weight in the two directions is different.Limit power w from advertisement to entity (v_a,v_e) the rank value that calculated by step 1 determines, embodies entity importance in document is introduced in advertisement；From entity to advertisement Limit power w (v_e,v_a) it is defined as the inverse i.e. 1/Out (v) of entity out-degree.Based on this figure, Random Walk Algorithm is used to find out rank It is worth Top-k1 the advertisement entity that k1 the highest node, i.e. user are liked.

The advertising user of game records was had for those, the game that it was played, use identical method, obtain this kind of Top-k2 game entity of advertising user preference.

3) for having the mobile advertising user of game records, log in the frequency of game according to mobile subscriber, build mobile The game preference set of advertising user.Specifically: for having the advertising user of game records, concentrate at game data, find out The game that each user played, by user's game login record game more than twice in a week, puts into the game set of preferences of user Close.Game sets of preferences, as a part for user characteristics, embodies user's preference on game selects.

4) for having the mobile advertising user of game records, in conjunction with entity preference and the game set of preferences of user of user Close, build the characteristic vector of user；For there is no the mobile advertising user of game records, entity preference based on user, build The characteristic vector of user.

For there being the advertising user of game records, by the advertisement entity set of user, game entity set and game Record merges, and portrays the preference profile of user from multiple dimensions, the final characteristic vector building reflection user's entirety preference, Dimension is user's game entity, advertisement entity and game records length sum.Form is as follows:

User characteristics={ { game entity 1, game entity 2 ... }, { advertisement entity 1, advertisement entity 2 ... }, { game Record 1, game records 2 ... } }

For not having the advertising user of game records, the advertisement entity set of user is used to build user characteristics vector, dimension Degree is the length of user advertising entity sets.Form is as follows:

User characteristics={ advertisement entity 1, advertisement entity 2 ... }.

5) for having the mobile advertising user of game records, in conjunction with game entity, advertisement entity, game sets of preferences three Individual part calculates the neighbour user of user；For there is no the mobile advertising user of game records, based on advertisement entity from having trip The mobile advertising user of play record finds neighbour user.

Calculate the neighbor lists of two kinds of different advertising user.For there being the advertising user of game records, user's is similar Degree calculating is divided into three parts: the similarity of game entity, the similarity of advertisement entity, the similarity of game records, often a part The similarity drawn is with different weights w₁,w₂,w₃Carry out aggregative weighted, calculate the final similarity of user.Wherein, w₁,w₂, w₃It is adjusted according to experiment needs, is initialized as 1/3,1/3,1/3, i.e. three part Similarity Measure result proportion phases With.Computing formula is as shown in 5-1:

Similar(u₁,u₂)=w₁*cosine(gameEntity₁,gameEntity₂)

+w₂*cosine(adEntity₁,adEntity₂)

+w₃*cosine(gameList₁,gameList₂) (5-1)

Wherein, gameEntity is game entity set, and adEntity is advertisement entity set, and gameList is that user swims Play set of records ends.Wherein cosine is cosine similarity, and computing formula is as shown in 5-2:

$\cos ine(u_i,u_j)=\frac{u_i*u_j}{\left|\right|u_i\left|\right|*\left|\right|u_j\left|\right|}=\frac{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x}_{i,k}*x_{j,k}}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{i,k}}^2}*\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{j,k}}^2}}---(5-2)$

Finally, according to the height of cosine similarity, Top-N the neighbour user of user is found out.

For not having the advertising user of game records, we are by the advertising user calculating this user with have game records Cosine similarity, build neighbour's user list of this user, shown in the following 5-3 of computing formula:

Similar(u₁,u₂)=cosine (adEntity₁,adEntity₂) (5-3)

Wherein, u₁It is the advertising user not having game records, u₂It is the advertising user having game records, final u₁Neighbour User is the advertising user having game records.

6) proposed algorithm based on User-Based provides the user with game recommdation list.

For each user, according to the game records in its neighbour user, below equation is used to recommend:

$Grade(u_i,{\mathrm{game}}_j)=\underset{v\∈Neigh\left(u_i\right)\∩{\mathrm{game}}_j\∈gameList\left(v\right)}{\mathrm{\Σ}}Similar(u_i,v)---(6-1)$

Wherein, Neigh (u_i) it is u_iTop-N neighbour user, gameList is user's game records；

Finally, according to the value of grade from big to small, front Top-K the game of user is found out as recommendation list.

Embodiment described above is only the preferred embodiments of the invention, not limits the practical range of the present invention with this, therefore The change that all shapes according to the present invention, principle are made, all should contain within the scope of the present invention.

Claims (6)

1. method is recommended in the mobile phone games towards mobile advertising user, it is characterised in that comprise the following steps:

1) carry out participle by advertisement and game being introduced content, utilize keyword extraction techniques based on figure, obtain advertisement The significant noun in content, i.e. entity is introduced with game；

2) for having the mobile advertising user of game records, build advertising aggregator and the game set of user, based on advertisement and The entity in content is introduced in game, builds the entity sets of preferences of user；For there is no the mobile advertising user of game records, structure Build the advertising aggregator of user, introduce the entity in content based on advertisement, build the entity sets of preferences of user；

3) for having the mobile advertising user of game records, log in the frequency of game according to mobile subscriber, build moving advertising The game preference set of user；

4) for having the mobile advertising user of game records, in conjunction with entity preference and the game sets of preferences of user of user, Build the characteristic vector of user；For there is no the mobile advertising user of game records, entity preference based on user, build user Characteristic vector；

5) for having the mobile advertising user of game records, in conjunction with game entity, advertisement entity, three portions of game sets of preferences Divide the neighbour user calculating user；For there is no the mobile advertising user of game records, based on advertisement entity from having game note The mobile advertising user of record is found neighbour user；

6) proposed algorithm based on User-Based provides the user with game recommdation list.

Method is recommended in mobile phone games towards mobile advertising user the most according to claim 1, it is characterised in that: in step 1) in, game is introduced data, use the conventional participle instrument with part-of-speech tagging to carry out participle, thus game is introduced Data resolve into the form of document-word；During participle, filter out non-nominal vocabulary；Fall into a trap operator node based on graph theory The method of importance degree calculates the importance of each entity node；Figure, figure is set up according to the syntopy of entity in entity sets The weight on middle limit is all 1, then uses the random walk restarted to calculate the rank value of each node, and formula is as shown in 1-1；

$rank\left(e_i\right)=(1-\mathrm{\α})+\mathrm{\α}*\underset{e_j\∈In\left(e_i\right)}{\Σ}rank\left(e_j\right)/\left|Out\left(e_j\right)\right|---\left(1-1\right)$

Wherein, rank (e_i) presentation-entity e_iSignificance level in a document, α represents the parameter that restarts of random walk, In (e_i) Refer to entity e_iNeighborhood, | Out (e_j) | refer to entity e_jNeighbours' number；

Finally, according to the rank value of each entity, extract each game and introduce the entity of the front N of data rank value；In like manner, it is possible to Use same method to calculate each advertisement and introduce the top n entity of data.

Method is recommended in mobile phone games towards mobile advertising user the most according to claim 1, it is characterised in that: in step 3), in, for having the advertising user of game records, concentrate at game data, find out the game that each user played, by user The game login record game more than twice in one week, puts into the game sets of preferences of user；Game sets of preferences is special as user The part levied, embodies user's preference on game selects.

Method is recommended in mobile phone games towards mobile advertising user the most according to claim 1, it is characterised in that: in step 4) in, for there being the advertising user of game records, by the advertisement entity set of user, game entity set and game records Merge, portray the preference profile of user from multiple dimensions, the final characteristic vector building reflection user's entirety preference, dimension Being user's game entity, advertisement entity and game records length sum, form is as follows:

User characteristics={ { game entity 1, game entity 2 ... }, { advertisement entity 1, advertisement entity 2 ... }, { game records 1, game records 2 ... } }

For not having the advertising user of game records, using the advertisement entity set of user to build user characteristics vector, dimension is The length of user advertising entity sets, form is as follows:

User characteristics={ advertisement entity 1, advertisement entity 2 ... }.

Method is recommended in mobile phone games towards mobile advertising user the most according to claim 1, it is characterised in that: in step 5), in, the neighbor lists of two kinds of different advertising user is calculated, for there being the advertising user of game records, the similarity meter of user Point counting is three parts: the similarity of game entity, the similarity of advertisement entity, the similarity of game records, and every part draws Similarity with different weights w₁,w₂,w₃Carry out aggregative weighted, calculate the final similarity of user；Wherein, w₁,w₂,w₃Root Testing needs factually to be adjusted, be initialized as 1/3,1/3,1/3, i.e. three part Similarity Measure result proportions are identical；Meter Calculate formula as shown in 5-1:

Similar(u₁,u₂)=w₁*cosine(gameEntity₁,gameEntity₂)

+w₂*cosine(adEntity₁,adEntity₂)

+w₃*cosine(gameList₁,gameList₂) (5-1)

Wherein, gameEntity is game entity set, and adEntity is advertisement entity set, and gameList is that user plays note Record set；Wherein cosine is cosine similarity, and computing formula is as shown in 5-2:

$\cos ine\left(u_i,u_j\right)=\frac{u_i*u_j}{\left|\right|u_i\left|\right|*\left|\right|u_j\left|\right|}=\frac{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x}_{i,k}*x_{j,k}}{\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{i,k}}^2}*\sqrt{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{x_{j,k}}^2}}---\left(5-2\right)$

Finally, according to the height of cosine similarity, Top-N the neighbour user of user is found out；

For there is no the advertising user of game records, the cosine phase of the advertising user by calculating this user and there is game records Like degree, build neighbour's user list of this user, shown in the following 5-3 of computing formula:

Similar(u₁,u₂)=cosine (adEntity₁,adEntity₂) (5-3)

Wherein, u₁It is the advertising user not having game records, u₂It is the advertising user having game records, final u₁Neighbour user It it is all the advertising user having game records.

Method is recommended in mobile phone games towards mobile advertising user the most according to claim 1, it is characterised in that: in step 6) in, for each user, according to the game records in its neighbour user, below equation is used to recommend:

$Grade(u_i,{\mathrm{game}}_j)=\underset{v\∈Neigh\left(u_i\right)\∩{\mathrm{game}}_j\∈gameList\left(v\right)}{\Σ}Similar(u_i,v)---\left(6-1\right)$

Wherein, Neigh (u_i) it is u_iTop-N neighbour user, gameList is user's game records；

Finally, according to the value of grade from big to small, front Top-K the game of user is found out as recommendation list.