CN105868298A

CN105868298A - Mobile phone game recommendation method based on binary decision tree

Info

Publication number: CN105868298A
Application number: CN201610172881.4A
Authority: CN
Inventors: 古万荣; 董守斌; 胡金龙; 付佳兵; 张铃启
Original assignee: South China University of Technology SCUT
Current assignee: South China University of Technology SCUT
Priority date: 2016-03-23
Filing date: 2016-03-23
Publication date: 2016-08-17

Abstract

The invention discloses a mobile phone game recommendation method based on a binary decision tree, which mainly determines whether a user is interested in a specific mobile game to be recommended by training a binary decision tree classification model, thereby reducing the influence of feature weights on recommendation from subjective assignment Transform to learn model to predict automatically. The feature of the present invention is to use a binary decision tree to determine the problem of feature selection and weight measurement in the game recommendation scene, and provide a corresponding recommendation model. Using the present invention can better ensure the accuracy of recommendation results and user preferences, and at the same time can ensure the diversity of recommendation results on the premise that no additional secondary selection is required for the recommendation results.

Description

A Mobile Game Recommendation Method Based on Binary Decision Tree

技术领域technical field

本发明涉及数据挖掘、用户行为预测领域，特别涉及一种基于二叉决策树的手机游戏推荐方法。The invention relates to the fields of data mining and user behavior prediction, in particular to a mobile game recommendation method based on a binary decision tree.

背景技术Background technique

在中国经济、互联网和娱乐业飞速发展的今天，越来越多的人注重业余休闲的质量。智能手机的发展，也带来了移动应用软件、移动广告和手机游戏业的飞速发展。由于游戏种类和适应层次的多样性，海量的手机游戏给用户也会带来信息过载问题，因此，对手机用户进行游戏推荐的方法也应运而生。Today, with the rapid development of China's economy, Internet and entertainment industry, more and more people pay attention to the quality of leisure time. The development of smart phones has also brought about the rapid development of mobile application software, mobile advertising and mobile game industry. Due to the diversity of game types and adaptation levels, a large number of mobile games will also bring information overload to users. Therefore, the method of game recommendation for mobile phone users has also emerged as the times require.

游戏推荐是推荐技术研究的拓展应用形式之一。海量的游戏催生了推荐的需求，推荐的预测结果同时也可以用于指导游戏开发商专注于有效的游戏开发，节省开发和广告成本。套用经典推荐技术核心原理，如基于内容的推荐和基于协同过滤的推荐会导致一些问题：1)基于内容的推荐容易让用户获得的游戏过于集中，不利于新游戏的推广。2)基于协同过滤的推荐不能及时更新用户偏好度模型。3)传统推荐方法没有细化考虑特定应用，如游戏推荐中的特征及其权重衡量问题。4)现有方法没有考虑到游戏推荐扩展问题。即推荐模型的普适性不强，不能适应推荐场景中推荐条件的变化。Game recommendation is one of the extended application forms of recommendation technology research. A large number of games have created a demand for recommendations, and the recommended prediction results can also be used to guide game developers to focus on effective game development, saving development and advertising costs. Applying the core principles of classic recommendation techniques, such as content-based recommendation and collaborative filtering-based recommendation, will lead to some problems: 1) Content-based recommendation tends to make the games that users get too concentrated, which is not conducive to the promotion of new games. 2) The recommendation based on collaborative filtering cannot update the user preference model in time. 3) Traditional recommendation methods do not consider specific applications in detail, such as features and their weight measurement in game recommendation. 4) Existing methods do not consider the game recommendation scaling problem. That is to say, the universality of the recommendation model is not strong, and it cannot adapt to the changes of the recommendation conditions in the recommendation scene.

发明内容Contents of the invention

本发明的目的在于克服现有技术的缺点与不足，提供一种基于二叉决策树的手机游戏推荐方法。The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a mobile game recommendation method based on a binary decision tree.

本发明的目的通过如下技术方案实现：一种基于二叉决策树的手机游戏推荐方法，具体包括以下步骤：The object of the present invention is achieved through the following technical solutions: a mobile game recommendation method based on a binary decision tree, specifically comprising the following steps:

S1、利用已有的用户玩游戏记录，进行预处理，确定近邻用户和远邻用户集合，获得用户、游戏、开发商、广告、地区、时段和分类；S1. Use the existing user game records to perform preprocessing, determine the set of nearby users and distant users, and obtain users, games, developers, advertisements, regions, time periods and classifications;

S2、根据步骤S1中获得的信息特征构建样本和特征值表；S2. Constructing samples and feature value tables according to the information features obtained in step S1;

S3、训练游戏推荐的二叉决策树模型；S3, training the binary decision tree model recommended by the game;

S4、根据训练模型，对用户进行游戏推荐，带推荐游戏可以是游戏的全体集合，也可以是经过预处理后的游戏子集。S4. According to the training model, recommend games to the user. The recommended games can be the entire set of games, or a subset of games after preprocessing.

所述步骤S1具体为：先构建“用户-游戏”二分图，然后使用基于用户方法确定近邻和远邻用户，设置参数由应用数据的规模决定，待推荐的关键特征根据应用场景确定。The step S1 is as follows: first construct the "user-game" bipartite graph, and then use the user-based method to determine the neighbors and distant neighbors. The setting parameters are determined by the scale of the application data, and the key features to be recommended are determined according to the application scenario.

在所述步骤S2的训练集中即近邻用户和远邻用户对应的内容，矩阵内容根据重合度计算，即当前用户和其列表中的用户在对应属性集中的重叠衡量，即考察集合的重叠问题，设某用户的特征值表为D_tr，公式如下：In the training set of step S2, that is, the content corresponding to the adjacent users and the distant users, the content of the matrix is calculated according to the coincidence degree, that is, the overlapping measurement of the current user and the users in the list in the corresponding attribute set, that is, the overlapping problem of the investigation set, Let the characteristic value table of a certain user be D _tr , the formula is as follows:

${D D.}_{t t r r} {| |}_{v v,, a a} = = \{\begin{matrix} 00,, & \frac{| | {a a}_{v v} | | \cap \cap | | {a a}_{u u} | |}{| | {a a}_{v v} | | \cup \cup | | {a a}_{u u} | |} < < T T h h r r e e s the s h h o o l l d d \\ 11,, & \frac{| | {a a}_{v v} | | \cap \cap | | {a a}_{u u} | |}{| | {a a}_{v v} | | \cup \cup | | {a a}_{u u} | |} &GreaterEqual; &Greater Equal; T T h h r r e e s the s h h o o l l d d \end{matrix}$

其中，v表示表中对应的某个k近邻用户或k远邻用户，a表示特征，u表示当前考察的用户。Among them, v represents a corresponding k-nearest neighbor user or k-distant neighbor user in the table, a represents a feature, and u represents the currently investigated user.

所述步骤S2中候选游戏中的矩阵元素的赋值由以下公式确定：The assignment of the matrix elements in the candidate game in the step S2 is determined by the following formula:

该公式表示，当游戏n对应的属性列a如果要赋值为1，则必须要其对应的属性a的值同时出现在游戏n和用户u玩过的游戏中。This formula indicates that if the attribute column a corresponding to game n is to be assigned a value of 1, the value of the corresponding attribute a must appear in both game n and the games played by user u.

所述步骤S3中构建二叉决策树关键在于其节点及其树枝的构建，节点层次越低，说明该节点特征越重要，使用信息增益比来表示特征节点的重要性，即信息增益比越大，则其节点层次越低。The key to constructing a binary decision tree in step S3 lies in the construction of its nodes and their branches. The lower the node level, the more important the characteristics of the node are. The information gain ratio is used to represent the importance of characteristic nodes, that is, the greater the information gain ratio , the lower the node level is.

所述步骤S3包括以下分步骤：The step S3 includes the following sub-steps:

S3.1基本定义：信息增益比即属性的信息增益与训练数据集的经验熵的比值：S3.1 Basic definition: The information gain ratio is the ratio of the information gain of the attribute to the experience entropy of the training data set:

${g g}_{R R} ((D D.,, a a)) = = \frac{g g ((D D.,, a a))}{H h ((D D.))}$

其中，H(D)表示训练数据集的经验熵，D即训练数据集，g(D,a)表示特征属性a在训练数据集D中的信息增益；Among them, H(D) represents the experience entropy of the training data set, D is the training data set, and g(D, a) represents the information gain of the feature attribute a in the training data set D;

数据集D的经验熵的计算公式如下：The calculation formula of the empirical entropy of the data set D is as follows:

$H h ((D D.)) = = - - {Σ Σ}_{k k = = 11}^{22} \frac{| | {c c}_{k k} | |}{| | D D. | |} {log log}_{22} \frac{| | {c c}_{k k} | |}{| | D D. | |},,$

其中，c_k表示的是类别，类别只有两种；Among them, c _k represents the category, and there are only two categories;

获得训练数据集整体的经验熵后，需要计算指定特征属性a在此训练数据集中的经验条件熵H(D|a)：After obtaining the overall experience entropy of the training data set, it is necessary to calculate the empirical conditional entropy H(D|a) of the specified feature attribute a in this training data set:

$H h ((D D. | | a a)) = = {Σ Σ}_{i i = = 11}^{n no} \frac{| | {D D.}_{i i} | |}{| | D D. | |} H h (({D D.}_{i i})) = = - - {Σ Σ}_{i i = = 11}^{n no} \frac{| | {D D.}_{i i} | |}{| | D D. | |} {Σ Σ}_{k k = = 11}^{22} \frac{| | {D D.}_{i i k k} | |}{| | {D D.}_{i i} | |} {log log}_{22} \frac{| | {D D.}_{i i k k} | |}{| | {D D.}_{i i} | |},,$

信息增益g(D,a)则是整体经验熵与经验条件熵的差：The information gain g(D,a) is the difference between the overall experience entropy and the experience conditional entropy:

g(D,a)＝H(D)-H(D|a)，g(D,a)=H(D)-H(D|a),

其中信息增益比是个相对概念，即当考察的集合对象变化时，需要重新计算，如构建节点的过程中，考察的集合已经在产生动态变化；Among them, the information gain ratio is a relative concept, that is, when the inspected set object changes, it needs to be recalculated, for example, in the process of building nodes, the inspected set is already undergoing dynamic changes;

步骤S3.2、模型训练：给定训练集D，即示例表中的k近邻用户和k远用户所对应的子表，和特征集A，其中A＝{手游开发商，广告，地区，时段，分类}，用户u的决策树Tr生成过程如下：Step S3.2, model training: given training set D, that is, the sub-table corresponding to k nearest neighbor users and k distant users in the example table, and feature set A, where A={mobile game developer, advertisement, region, Period, classification}, the process of generating the decision tree Tr of user u is as follows:

1)若训练集D中所有记录都属于同一个类(0或1)，则Tr为单节点二叉树，并将该类设置为该节点值，返回Tr；1) If all the records in the training set D belong to the same class (0 or 1), then Tr is a single-node binary tree, and this class is set as the node value, and Tr is returned;

2)如果特征集A为空集，则将训练集D中具有最多记录的类别作为该节点的类别，返回Tr；2) If the feature set A is an empty set, then use the category with the most records in the training set D as the category of the node, and return Tr;

3)否则，计算剩余特征集各特征的信息增益比g_R(D,a)，并选择出最大信息增益比对应的特征a；3) Otherwise, calculate the information gain ratio g _R (D,a) of each feature in the remaining feature set, and select the feature a corresponding to the maximum information gain ratio;

4)对选出来的最大特征a的每一个可能值(0或1)将D切割为两个非空子集D₀和D₁，将子集D₀和D₁中记录数最多的类作为标记，构建子节点，由该节点及其子节点，返回供递归使用的Tr；4) For each possible value (0 or 1) of the selected maximum feature a, cut D into two non-empty subsets D ₀ and D ₁ , and use the class with the largest number of records in the subsets D ₀ and D ₁ as a mark , build a child node, and return Tr for recursive use from this node and its child nodes;

5)以非空子集D₀和D₁为训练集，以A\{a}为新的特征集，递归调用步骤1～步骤4，得到子树Tr₀和Tr₁，并返回相应子树。5) Take the non-empty subsets D ₀ and D ₁ as the training set and A\{a} as the new feature set, recursively call steps 1 to 4 to obtain subtrees Tr ₀ and Tr ₁ , and return the corresponding subtrees.

本发明与现有技术相比，具有如下优点和有益效果：Compared with the prior art, the present invention has the following advantages and beneficial effects:

1、本发明可以有效地融合基于内容和基于协同过滤的基本思想，使推荐趋于融合推荐模型，有利于提高推荐准确度。1. The present invention can effectively integrate the basic ideas of content-based and collaborative filtering, so that the recommendation tends to integrate the recommendation model, which is conducive to improving the accuracy of recommendation.

2、本发明对游戏推荐中的各种特征进行重要度学习，而无需依赖于经验权重，有利于推荐结果的客观公正。2. The present invention learns the importance of various features in game recommendation without relying on experience weights, which is conducive to the objective and fair recommendation results.

3、本发明区别于以往的推荐方法，不容易产生基于内容的推荐游戏聚合，即较大限度地确保了推荐游戏的多样化。3. The present invention is different from previous recommendation methods, and it is not easy to generate content-based aggregation of recommended games, that is, to ensure the diversification of recommended games to a large extent.

4、本发明的方法容易扩展，如特征可以根据不同的游戏应用需求进行增减，稍加修改还可以扩展到其他领域，如音乐推荐、广告推荐、应用推荐等。4. The method of the present invention is easy to expand. For example, features can be increased or decreased according to different game application requirements, and can be extended to other fields with slight modifications, such as music recommendation, advertisement recommendation, application recommendation, etc.

附图说明Description of drawings

图1为本发明所述的基于二叉决策树的手机游戏推荐方法的处理步骤流程图。Fig. 1 is a flow chart of the processing steps of the binary decision tree-based mobile game recommendation method of the present invention.

具体实施方式detailed description

下面结合实施例及附图对本发明作进一步详细的描述，但本发明的实施方式不限于此。The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

本发明具体做法步骤如下，如图1所示：The concrete practice step of the present invention is as follows, as shown in Figure 1:

步骤S1、预处理。Step S1, preprocessing.

该步骤首先要确定近邻用户和远邻用户集合。在本方法中，先构建“用户-游戏”二分图，然后使用基于用户(User-based)方法确定若干近邻和远邻用户，这里可以设置参数为m，该参数由应用数据的规模决定，一般将其设置为总用户数的10％。待推荐的关键特征根据应用场景确定，如应用场景中的游戏数据有“开发商”，则将其作为特征之一，以此类推。In this step, the set of near-neighbor users and far-neighbor users must be determined first. In this method, first construct the "user-game" bipartite graph, and then use the user-based (User-based) method to determine a number of nearby and distant users. Here, the parameter can be set to m, which is determined by the scale of the application data. Generally, Set it to 10% of the total number of users. The key features to be recommended are determined according to the application scenario. If the game data in the application scenario has "developer", it will be used as one of the features, and so on.

这些预处理产生的结果将为模型学习提供基本元素。The results produced by these preprocessing will provide the basic elements for model learning.

步骤S2、模型学习准备。Step S2, preparing for model learning.

该步骤主要是构建用户的样本和特征值表，如表1所示，这里设某用户的特征值表为D_tr。This step is mainly to construct user samples and feature value tables, as shown in Table 1, where the feature value table of a certain user is set as D _tr .

表1 D_tr表结构示例Table 1 _Dtr table structure example

(备注：表中行表示近邻用户u_i、远邻用户u_i′以及待推荐游戏n_i；列表示关联特征以及最后一列的是否接受的标记。表中的元素表示与当前用户是否有重合之处，如有重合，则元素为1，否则为0。)(Remarks: The row in the table represents the nearby user u _i , the distant user u _i ′, and the game to be recommended n _i ; the column represents the associated feature and the mark of acceptance in the last column. The elements in the table represent whether there is overlap with the current user , if there is overlap, the element is 1, otherwise it is 0.)

表1中涵盖了决策树学习的训练集和待预测集，其中，训练集由k近邻用户和k远用户构成，待预测集即本节需要确认该用户是否接受待推荐游戏的构成。该表有一些特点：Table 1 covers the training set and the to-be-predicted set of decision tree learning, where the training set consists of k nearest neighbor users and k distant users, and the to-be-predicted set is the composition that needs to be confirmed in this section whether the user accepts the game to be recommended. The table has some characteristics:

1)k近邻用户对应的“接受”字段全部为1。1) The "accept" fields corresponding to the k-nearest neighbor users are all 1.

2)k远用户对应的“接受”字段全部为0。2) The "Accept" field corresponding to the k-distant user is all 0.

3)待预测集的“接受”字段未知，即算法需要填补的字段。3) The "acceptance" field of the set to be predicted is unknown, that is, the field that the algorithm needs to fill.

4)在训练集中(k近邻用户和k远用户对应的内容)，矩阵内容根据重合度计算，即当前用户和其列表中的用户在对应属性集中的重叠衡量，即考察集合的重叠问题，公式如下：4) In the training set (the content corresponding to k nearest neighbor users and k distant users), the matrix content is calculated according to the coincidence degree, that is, the overlap measurement between the current user and the users in its list in the corresponding attribute set, that is, to investigate the overlapping problem of the set, the formula as follows:

${D D.}_{t t r r} {| |}_{v v,, a a} = = \{\begin{matrix} 00,, & \frac{| | {a a}_{v v} | | \cap \cap | | {a a}_{u u} | |}{| | {a a}_{v v} | | \cup \cup | | {a a}_{u u} | |} < < T T h h r r e e s the s h h o o l l d d \\ 11,, & \frac{| | {a a}_{v v} | | \cap \cap | | {a a}_{u u} | |}{| | {a a}_{v v} | | \cup \cup | | {a a}_{u u} | |} &GreaterEqual; &Greater Equal; T T h h r r e e s the s h h o o l l d d \end{matrix} - - - - - - ((11))$

其中，k表示表中对应的某个k近邻用户或k远用户，a表示特征(即开发商、广告、地区、时段或分类)，u表示当前考察的用户。该公式表示，当用户u和用户v在某个属性中的集合重叠度大于等于阈值T时，对应矩阵元素赋值为1，否则赋值为0。Among them, k represents a corresponding k-nearest neighbor user or k-distant user in the table, a represents a feature (that is, developer, advertisement, region, time period or classification), and u represents the currently investigated user. This formula indicates that when the set overlap between user u and user v in a certain attribute is greater than or equal to the threshold T, the corresponding matrix element is assigned a value of 1, otherwise the value is assigned 0.

5)候选游戏中的矩阵元素的赋值由以下公式确定：5) The assignment of matrix elements in the candidate game is determined by the following formula:

该公式表示，当游戏n对应的属性列a如果要赋值为1，则必须要其对应的属性a的值同时出现在游戏n和用户u玩过的游戏中。例如，待推荐游戏n的开发商出现了“网易”，且当前用户u玩过的游戏也有同样的开发商关联，则对应的矩阵元素D_tr|_n,a赋值为1。This formula indicates that if the attribute column a corresponding to game n is to be assigned a value of 1, the value of the corresponding attribute a must appear in both game n and the games played by user u. For example, if "NetEase" appears as the developer of the game n to be recommended, and the game played by the current user u is also associated with the same developer, then the corresponding matrix element D _tr | _n,a is assigned a value of 1.

步骤S3、二叉决策树学习Step S3, binary decision tree learning

构建二叉决策树关键在于其节点及其树枝的构建，节点层次越低，说明该节点特征越重要，本方法使用信息增益比来表示特征节点的重要性，即信息增益比越大，则其节点层次越低。The key to constructing a binary decision tree lies in the construction of its nodes and their branches. The lower the node level, the more important the characteristics of the node are. This method uses the information gain ratio to represent the importance of feature nodes, that is, the greater the information gain ratio, the more The lower the node level.

步骤S3.1基本定义Step S3.1 Basic Definition

信息增益比即属性的信息增益与训练数据集的经验熵的比值：The information gain ratio is the ratio of the information gain of the attribute to the experience entropy of the training data set:

${g g}_{R R} ((D D.,, a a)) = = \frac{g g ((D D.,, a a))}{H h ((D D.))} - - - - - - ((33))$

其中，H(D)表示训练数据集的经验熵，D即训练数据集，g(D,a)表示特征属性a在训练数据集D中的信息增益。数据集D的经验熵的计算公式如下：Among them, H(D) represents the experience entropy of the training data set, D is the training data set, and g(D, a) represents the information gain of the feature attribute a in the training data set D. The calculation formula of the empirical entropy of the data set D is as follows:

$H h ((D D.)) = = - - {Σ Σ}_{k k = = 11}^{22} \frac{| | {c c}_{k k} | |}{| | D D. | |} {log log}_{22} \frac{| | {c c}_{k k} | |}{| | D D. | |} - - - - - - ((44))$

其中，c_k表示的是类别，在小节提出的方法中，类别只有两种。获得训练数据集整体的经验熵后，需要计算指定特征属性a在此训练数据集中的经验条件熵H(D|a)：Among them, c _k represents the category, and in the method proposed in the subsection, there are only two categories. After obtaining the overall experience entropy of the training data set, it is necessary to calculate the empirical conditional entropy H(D|a) of the specified feature attribute a in this training data set:

$H h ((D D. | | a a)) = = {Σ Σ}_{i i = = 11}^{n no} \frac{| | {D D.}_{i i} | |}{| | D D. | |} H h (({D D.}_{i i})) = = - - {Σ Σ}_{i i = = 11}^{n no} \frac{| | {D D.}_{i i} | |}{| | D D. | |} {Σ Σ}_{k k = = 11}^{22} \frac{| | {D D.}_{i i k k} | |}{| | {D D.}_{i i} | |} {log log}_{22} \frac{| | {D D.}_{i i k k} | |}{| | {D D.}_{i i} | |} - - - - - - ((55))$

g(D,a)＝HkD)-H(D|a) (6)g(D,a)=HkD)-H(D|a) (6)

这里需要注意的是，信息增益比是个相对概念，即当考察的集合对象变化时，需要重新计算，如构建节点的过程中，考察的集合已经在产生动态变化。It should be noted here that the information gain ratio is a relative concept, that is, when the set object under investigation changes, it needs to be recalculated. For example, during the process of building nodes, the set under investigation is already undergoing dynamic changes.

步骤S3.2、模型训练Step S3.2, model training

给定训练集D(即示例表中的k近邻用户和k远用户所对应的子表)和特征集A，(A＝{手游开发商，广告，地区，时段，分类})，用户u的决策树Tr生成过程如下：Given a training set D (that is, the sub-table corresponding to k nearest neighbor users and k distant users in the example table) and feature set A, (A={mobile game developer, advertisement, region, time period, classification}), user u The decision tree Tr generation process is as follows:

若训练集D中所有记录都属于同一个类(0或1)，则Tr为单节点二叉树，并将该类设置为该节点值，返回Tr。If all records in the training set D belong to the same class (0 or 1), then Tr is a single-node binary tree, and this class is set as the node value, and Tr is returned.

2)如果特征集A为空集，则将训练集D中具有最多记录的类别作为该节点的类别，返回Tr。2) If the feature set A is an empty set, then use the category with the most records in the training set D as the category of the node, and return Tr.

3)否则，计算剩余特征集各特征的信息增益比g_R(D,a)，并选择出最大信息增益比对应的特征a。3) Otherwise, calculate the information gain ratio g _R (D,a) of each feature in the remaining feature set, and select the feature a corresponding to the maximum information gain ratio.

4)对选出来的最大特征a的每一个可能值(0或1)将D切割为两个非空子集D₀和D₁，将子集D₀和D₁中记录数最多的类作为标记，构建子节点，由该节点及其子节点，返回供递归使用的Tr。4) For each possible value (0 or 1) of the selected maximum feature a, cut D into two non-empty subsets D ₀ and D ₁ , and use the class with the largest number of records in the subsets D ₀ and D ₁ as a mark , build a child node, and return Tr for recursive use from this node and its child nodes.

步骤S4、游戏推荐。Step S4, game recommendation.

根据训练模型，对用户进行游戏推荐，带推荐游戏可以是游戏的全体集合，也可以是经过预处理后的游戏子集。According to the training model, game recommendations are made to users. The recommended games can be the entire collection of games, or a subset of games after preprocessing.

(1)根据本发明的步骤S1，利用已有的用户玩游戏记录，进行预处理，获得用户、游戏、开发商、广告、地区、时段和分类。(1) According to step S1 of the present invention, use existing user game play records for preprocessing to obtain users, games, developers, advertisements, regions, time periods and classifications.

(2)根据本发明的步骤S2，利用这些特征构建样本和特征值表。(2) According to step S2 of the present invention, use these features to construct samples and feature value tables.

(3)根据本发明的步骤S3，训练游戏推荐的二叉决策树模型。(3) According to step S3 of the present invention, train the binary decision tree model of game recommendation.

(4)根据本发明的步骤S4，利用二叉决策树模型给用户推荐游戏。(4) According to step S4 of the present invention, use the binary decision tree model to recommend games to the user.

上述实施例为本发明较佳的实施方式，但本发明的实施方式并不受上述实施例的限制，其他的任何未背离本发明的精神实质与原理下所作的改变、修饰、替代、组合、简化，均应为等效的置换方式，都包含在本发明的保护范围之内。The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims

1. method is recommended in mobile phone games based on binary decision tree, it is characterised in that comprise the following steps:

S1, utilize existing user to play game records, pre-process, determine neighbour user and remote adjacent user's set, it is thus achieved that use Family, game, developer, advertisement, area, period and classification；

S2, build sample and the list of feature values according to the information characteristics obtained in step S1；

The Binary decision tree-model that S3, training game are recommended；

S4, according to training pattern, user is carried out game recommdation, band recommended games can be all set of game, it is also possible to It is through pretreated game subset.

Method is recommended in mobile phone games based on binary decision tree the most according to claim 1, it is characterised in that: described step S1, particularly as follows: first build " user-game " bipartite graph, then uses and determines neighbour and remote adjacent user based on user method, if Putting parameter to be determined by the scale applying data, key feature to be recommended determines according to application scenarios.

Method is recommended in mobile phone games based on binary decision tree the most according to claim 1, it is characterised in that: in described step The content that in the training set of rapid S2, i.e. neighbour user is corresponding with remote adjacent user, matrix content calculates according to registration, i.e. active user With the overlapping measurement in corresponding property set of the user in its list, i.e. investigate the overlap problem of set, if the characteristic value of certain user Table is D_tr, formula is as follows:

D_{t r} |_{v, a} = \{\begin{matrix} 0, & \frac{| a_{v} | \cap | a_{u} |}{| a_{v} | \cup | a_{u} |} < T h r e s h o l d \\ 1, & \frac{| a_{v} | \cap | a_{u} |}{| a_{v} | \cup | a_{u} |} &GreaterEqual; T h r e s h o l d \end{matrix}

Wherein, v represents the far adjacent user of certain k neighbour user corresponding in table or k, and a represents that feature, u represent the use of current investigation Family.

Method is recommended in mobile phone games based on binary decision tree the most according to claim 1, it is characterised in that: described step In S2, the assignment of the matrix element in candidate's game is determined by below equation:

This formula represents, if when game attribute column a corresponding to n to be entered as 1, then have to its correspondence attribute a value simultaneously Occur in the game that game n and user u played.

Method is recommended in mobile phone games based on binary decision tree the most according to claim 1, it is characterised in that: described step Building binary decision tree in S3 and it is critical only that the structure of its node and branch thereof, node level is the lowest, illustrates that this node diagnostic is the heaviest Want, use information gain ratio to represent the importance of characteristic node, i.e. information gain ratio is the biggest, then its node level is the lowest.

Method is recommended in mobile phone games based on binary decision tree the most according to claim 1, it is characterised in that: described step S3 include following step by step:

S3.1 basic definition: the information gain of the information gain i.e. attribute of ratio and the ratio of the empirical entropy of training dataset:

g_{R} (D, a) = \frac{g (D, a)}{H (D)}

Wherein, H (D) represents the empirical entropy of training dataset, D i.e. training dataset, and (D a) represents that characteristic attribute a is in training to g Information gain in data set D；

The computing formula of the empirical entropy of data set D is as follows:

H (D) = - Σ_{k = 1}^{2} \frac{| c_{k} |}{| D |} \log_{2} \frac{| c_{k} |}{| D |},

Wherein, c_kRepresent is classification, and classification only has two kinds；

After obtaining the empirical entropy that training dataset is overall, need to calculate the empirical condition that specific characteristic attribute a concentrates at this training data Entropy H (D | a):

H (D | a) = Σ_{i = 1}^{n} \frac{| D_{i} |}{| D |} H (D_{i}) = - Σ_{i = 1}^{n} \frac{| D_{i} |}{| D |} Σ_{k = 1}^{2} \frac{| D_{i k} |}{| D_{i} |} \log_{2} \frac{| D_{i k} |}{| D_{i} |},

Information gain g (D, a) is then the difference of overall experience entropy and empirical condition entropy:

G (D, a)=H (D)-H (D | a),

Wherein information gain ratio is a relative concept, i.e. when the collection object change investigated, needs to recalculate, as built joint During Dian, the set of investigation is producing dynamically change；

Step S3.2, model training: the k neighbour user in given training set D, i.e. sample table and the sublist corresponding to the remote user of k, With feature set A, wherein A={ hand trip developer, advertisement, area, period, classification }, the decision tree Tr of user u generates process such as Under:

1) if all records broadly fall into same class (0 or 1) in training set D, then Tr is single node binary tree, and such is set It is set to this nodal value, returns Tr；

2) if feature set A is empty set, then using training set D having the classification at most recorded as the classification of this node, Tr is returned；

3) information gain otherwise, calculating each feature of residue character collection compares g_R(D, a), and it is corresponding to select the maximum information ratio of gains Feature a；

4) D is cut into by each probable value (0 or 1) of maximum feature a elected two nonvoid subsets D₀And D₁, will Subset D₀And D₁The most class of middle record number, as mark, builds child node, by this node and child node thereof, returns and makes for recurrence Tr；

5) with nonvoid subset D₀And D₁For training set, with A { a} is new feature set, and recursive call step 1～step 4 obtain Subtree Tr₀And Tr₁, and return corresponding subtree.