CN111882381A - Travel recommendation method based on collaborative memory network - Google Patents
Travel recommendation method based on collaborative memory network Download PDFInfo
- Publication number
- CN111882381A CN111882381A CN202010618323.2A CN202010618323A CN111882381A CN 111882381 A CN111882381 A CN 111882381A CN 202010618323 A CN202010618323 A CN 202010618323A CN 111882381 A CN111882381 A CN 111882381A
- Authority
- CN
- China
- Prior art keywords
- user
- spot
- scenic
- scenic spot
- vector
- 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
- 238000000034 method Methods 0.000 title claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 12
- 230000007246 mechanism Effects 0.000 claims abstract description 8
- 239000013598 vector Substances 0.000 claims description 127
- 239000011159 matrix material Substances 0.000 claims description 90
- 238000010606 normalization Methods 0.000 claims description 4
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims 1
- 238000012163 sequencing technique Methods 0.000 claims 1
- 238000001914 filtration Methods 0.000 abstract description 8
- 230000001568 sexual effect Effects 0.000 abstract 1
- 238000013135 deep learning Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 230000000717 retained effect Effects 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0631—Item recommendations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/14—Travel agencies
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Tourism & Hospitality (AREA)
- Economics (AREA)
- Marketing (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Human Resources & Organizations (AREA)
- Primary Health Care (AREA)
- Development Economics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
本发明公开一种基于协同记忆网络的旅游推荐方法，首先使用神经网络将用户与景点进行向量化，然后引入了记忆网络，在记忆网络中采用了注意力机制的方法，最后在预测用户对景点的得分时，将隐因子模型的全局结构信息与记忆网络的局部邻域结构信息进行结合。本发明同时考虑用户的局部邻域结构信息以及景点的局部邻域结构信息，并与隐因子模型的协同过滤方法进行融合，实现了一种基于协同记忆网络的旅游推荐方法，以此达到高准确性和高个性化旅游推荐的目标。The invention discloses a travel recommendation method based on a collaborative memory network. First, a neural network is used to vectorize users and scenic spots, then a memory network is introduced, and an attention mechanism method is adopted in the memory network. When the score is obtained, the global structure information of the latent factor model is combined with the local neighborhood structure information of the memory network. The invention simultaneously considers the local neighborhood structure information of the user and the local neighborhood structure information of the scenic spot, and integrates it with the collaborative filtering method of the latent factor model to realize a travel recommendation method based on a collaborative memory network, so as to achieve high accuracy The goal of sexual and highly personalized travel recommendations.
Description
技术领域technical field
本发明涉及智能推荐技术领域，具体涉及一种基于协同记忆网络的旅游推荐方法。The invention relates to the technical field of intelligent recommendation, in particular to a travel recommendation method based on a collaborative memory network.
背景技术Background technique
随着社会的发展和人们生活水平的提高，越来越多的人选择外出旅游，由于近年来旅游业的蓬勃发展以及互联网技术的迅速普及，使得目前一些主流旅游信息服务平台的旅游信息过载，人们对于旅游景点的个性化推荐需求变得越来越大，能在人们外出旅游时推荐合适的景点显得尤为重要。With the development of society and the improvement of people's living standards, more and more people choose to travel abroad. Due to the vigorous development of tourism and the rapid popularization of Internet technology in recent years, the current tourism information overload of some mainstream tourism information service platforms, People's demand for personalized recommendation of tourist attractions has become larger and larger, and it is particularly important to be able to recommend suitable attractions when people travel.
传统的基于深度学习的旅游推荐方法，主要利用基于隐因子模型的协同过滤方法来构建深度学习框架，并以此来进行推荐。这些推荐方法仅考虑了隐因子模型的全局结构信息，而没有考虑到局部邻域结构信息，这会造成的景点推荐结果准确性不高和个性化不足等问题。The traditional deep learning-based travel recommendation method mainly uses the collaborative filtering method based on the latent factor model to build a deep learning framework, and uses this to make recommendations. These recommendation methods only consider the global structure information of the latent factor model, but do not consider the local neighborhood structure information, which will cause problems such as low accuracy and insufficient personalization of scenic spot recommendation results.
发明内容SUMMARY OF THE INVENTION
本发明所要解决的是现有基于深度学习的旅游推荐方法存在景点推荐结果准确性不高和个性化不足的问题，提供一种基于协同记忆网络的旅游推荐方法。The invention aims to solve the problems of low accuracy of scenic spot recommendation results and insufficient personalization in the existing deep learning-based travel recommendation methods, and provides a travel recommendation method based on a collaborative memory network.
为解决上述问题，本发明是通过以下技术方案实现的：In order to solve the above-mentioned problems, the present invention is achieved through the following technical solutions:
一种基于协同记忆网络的旅游推荐方法，包括步骤如下：A travel recommendation method based on a collaborative memory network, comprising the following steps:
步骤1、利用爬虫工具从旅游网站上采集用户的游记，从中抽取出各个用户所访问过的景点，并对所有用户及其所有景点进行分别编号；Step 1. Use the crawler tool to collect the travel notes of the user from the travel website, extract the scenic spots visited by each user, and number all the users and all the scenic spots respectively;
步骤2、将每个用户访问过的所有景点都分别作为一个集合，构建关于用户的景点邻域集；同时，将访问过每个景点的所有用户都分别作为一个集合，构建关于景点的用户邻域集；Step 2. Take all the scenic spots visited by each user as a set, and construct a neighborhood set of scenic spots about the user; at the same time, take all the users who have visited each scenic spot as a set, and construct the user neighborhood of scenic spots. domain set;
步骤3、利用神经网络将所有用户和所有景点分别映射到两个特征向量空间中，其中一个特征向量空间为用户向量矩阵和景点向量矩阵，另一个特征向量空间为用户外部记忆矩阵和景点外部记忆矩阵；Step 3. Use neural network to map all users and all scenic spots into two feature vector spaces, one of which is the user vector matrix and the scenic spot vector matrix, and the other feature vector space is the user external memory matrix and the scenic spot external memory. matrix;
步骤4、对于每个用户u，基于步骤3所得到的用户向量矩阵和景点向量矩阵，计算用户u与关于景点i的用户邻域集中各个用户v的相似度quiv，并对用户u与其用户邻域集中各个用户v的相似度quiv进行归一化处理后得到用户u的权重系数puiv；其中：Step 4. For each user u, based on the user vector matrix and the scenic spot vector matrix obtained in step 3, calculate the similarity qiv between the user u and each user v in the user neighborhood set about the scenic spot i, and compare the user u with its users. The similarity q uiv of each user v in the neighborhood set is normalized to obtain the weight coefficient p uiv of user u; where:
步骤5、将用户u的权重系数puiv作为注意力机制中的权重，并将关于景点i的用户邻域集中各个用户v在用户外部记忆矩阵中的用户特征向量进行加权求和，得到关于用户u的邻域结构信息oui；Step 5. Use the weight coefficient p uiv of the user u as the weight in the attention mechanism, and perform a weighted summation of the user feature vectors of each user v in the user's external memory matrix in the user neighborhood set about the scenic spot i, to obtain the information about the user. u's neighborhood structure information o ui ;
步骤6、对于每个景点i，基于步骤3所得到的用户向量矩阵和景点向量矩阵，计算景点i与关于用户u的景点邻域集中各个景点j的相似度qiuj，并对景点i与其景点邻域集中各个景点j的相似度qiuj进行归一化处理后得到景点i的权重系数piuj；其中：Step 6. For each scenic spot i, based on the user vector matrix and the scenic spot vector matrix obtained in step 3, calculate the similarity q iuj between the scenic spot i and each scenic spot j in the scenic spot neighborhood set about the user u, and compare the scenic spot i and its scenic spots. The similarity q iuj of each scenic spot j in the neighborhood set is normalized to obtain the weight coefficient p iuj of the scenic spot i; wherein:
步骤7、将景点i的权重系数piuj作为注意力机制中的权重，并将用户u的景点领域集中各个景点j在景点外部记忆矩阵中的景点特征向量进行加权求和，得到关于景点i的邻域结构信息oiu；Step 7. Take the weight coefficient p iuj of the scenic spot i as the weight in the attention mechanism, and perform the weighted summation of the scenic spot feature vectors of each scenic spot j in the scenic spot external memory matrix in the scenic spot field of user u to obtain the information about the scenic spot i. neighborhood structure information o iu ;
步骤8、将用户u在用户向量矩阵中的用户特征向量与景点i在景点向量矩阵中的景点特征向量进行元素积，得到用户u对景点i的全局结构信息；Step 8. Perform element product of the user feature vector of user u in the user vector matrix and the scenic spot feature vector of scenic spot i in the scenic spot vector matrix to obtain the global structure information of user u to scenic spot i;
步骤9、将步骤5所得到的关于用户u的邻域结构信息oui、步骤7所得到的关于景点i的邻域结构信息oiu、以及步骤8所得到的用户u对景点i的全局结构信息进行拼接，并将拼接后的向量输入到多层神经网络中去得到用户u对景点i的最终评分；Step 9. Combine the neighborhood structure information o ui of user u obtained in step 5, the neighborhood structure information o iu of scenic spot i obtained in step 7, and the global structure of user u to scenic spot i obtained in step 8. The information is spliced, and the spliced vector is input into the multi-layer neural network to obtain the final score of the scenic spot i by user u;
步骤10、当需要针对某个目标用户推荐景点时，则将该目标用户对各个景点的最终评分进行排序，并选出排名靠前的K个景点，即得到关于该目标用户的Top-K旅游景点推荐；Step 10. When it is necessary to recommend scenic spots for a target user, sort the final scores of each scenic spot by the target user, and select the top K scenic spots, that is, get the Top-K tourism about the target user. recommended places;
上述，mu表示用户u在用户向量矩阵的用户特征向量，mv表示用户v在用户向量矩阵中的用户特征向量；ei表示景点i在景点向量矩阵中的景点特征向量，ej表示景点j在景点向量矩阵中的景点特征向量；N(i)表示所有访问过景点i的用户所构成的关于景点i的用户邻域集，S(u)表示用户u访问过的所有景点所构成的关于用户u的景点邻域集；u＝1,2,…,n，n表示用户的个数；i＝1,2,…,m，m表示景点的个数；其中K为设定的推荐景点的个数。In the above, m u represents the user feature vector of user u in the user vector matrix, m v represents the user feature vector of user v in the user vector matrix; e i represents the scenic spot feature vector of scenic spot i in the scenic spot vector matrix, and e j represents the scenic spot j is the scenic spot feature vector in the scenic spot vector matrix; N(i) represents the user neighborhood set about the scenic spot i formed by all the users who have visited the scenic spot i, and S(u) represents all the scenic spots that the user u has visited. About the scenic spot neighborhood set of user u; u = 1, 2, ..., n, n represents the number of users; i = 1, 2, ..., m, m represents the number of scenic spots; where K is the set recommendation number of attractions.
上述步骤3中，用户向量矩阵和用户外部记忆矩阵的维度均为n×d，景点向量矩阵和景点外部记忆矩阵的维度均为m×d，其中n表示用户的个数，m表示景点的个数，d表示矩阵的维度。In the above step 3, the dimensions of the user vector matrix and the user external memory matrix are both n × d, and the dimensions of the scenic spot vector matrix and the external memory matrix of the scenic spot are both m × d, where n represents the number of users, and m represents the number of scenic spots. number, d represents the dimension of the matrix.
上述步骤3中，还进一步包括如下过程：使用广义矩阵分解方法对用户向量矩阵和景点向量矩阵进行预训练更新。In the above step 3, the following process is further included: using the generalized matrix decomposition method to pre-train and update the user vector matrix and the sights vector matrix.
上述步骤5中，关于用户u的邻域结构信息oui为：In the above step 5, the neighborhood structure information o ui of user u is:
式中，puiv表示用户u的权重系数，cv表示关于景点i的用户邻域集中各个用户v在用户外部记忆矩阵中的用户特征向量，N(i)表示所有访问过景点i的用户所构成的用户邻域集，u＝1,2,…,n，n表示用户的个数；i＝1,2,…,m，m表示景点的个数。In the formula, p uiv represents the weight coefficient of user u, cv represents the user feature vector of each user v in the user's external memory matrix in the user neighborhood set about scenic spot i, and N(i) represents all users who have visited scenic spot i. The formed user neighborhood set, u=1,2,...,n, n represents the number of users; i=1,2,...,m, m represents the number of scenic spots.
上述步骤7中，关于景点i的邻域结构信息oiu为：In the above step 7, the neighborhood structure information o iu about the scenic spot i is:
式中，piuj表示景点i的权重系数，yj表示关于用户u的景点邻域集中各个景点j在景点外部记忆矩阵中的景点特征向量，S(u)表示用户u访问过的所有景点所构成的景点邻域集，u＝1,2,…,n，n表示用户的个数；i＝1,2,…,m，m表示景点的个数。In the formula, p iuj represents the weight coefficient of the scenic spot i, y j represents the scenic spot feature vector of each scenic spot j in the external memory matrix of the scenic spot in the scenic spot neighborhood set about the user u, and S(u) represents all the scenic spots visited by the user u. The formed scenic spot neighborhood set, u = 1, 2, ..., n, n represents the number of users; i = 1, 2, ..., m, m represents the number of scenic spots.
本发明通过利用基于邻域或记忆的协同过滤方法，同时考虑用户的局部邻域结构信息以及景点的局部邻域结构信息，并与隐因子模型的协同过滤方法进行融合，实现了一种基于协同记忆网络的旅游推荐方法，以此达到高准确性和高个性化旅游推荐的目标。The present invention realizes a collaborative filtering method based on collaborative filtering by utilizing the collaborative filtering method based on neighborhood or memory, taking into account the local neighborhood structure information of the user and the local neighborhood structure information of scenic spots, and integrating with the collaborative filtering method of the latent factor model. A travel recommendation method based on memory network, so as to achieve the goal of high accuracy and high personalized travel recommendation.
与现有技术相比，本发明具有如下特点：Compared with the prior art, the present invention has the following characteristics:
1、本发明使用神经网络将用户与景点进行向量化，即将用户和景点映射到低维度向量空间中，通过向量来表示。该方法将自然语言表示为简单的向量，不仅保持了原来数据的含义，还极大地简化了计算，使得用户与景点的表示更加准确、合理地与旅游推荐进行结合；1. The present invention uses a neural network to vectorize users and scenic spots, that is, map users and scenic spots into a low-dimensional vector space, and represent them by vectors. This method represents natural language as a simple vector, which not only maintains the meaning of the original data, but also greatly simplifies the calculation, making the representation of users and attractions more accurate and reasonable to combine with travel recommendation;
2、本发明引入了记忆网络，在记忆网络中采用了注意力机制的方法，既考虑了用户与其邻域内用户的相似性，又考虑了景点与其邻域内景点的相似性，得到了包含用户邻域以及景点邻域的局部邻域结构信息，从而能够增强旅游推荐的个性化和准确性；2. The present invention introduces a memory network, and adopts the method of attention mechanism in the memory network, which not only considers the similarity between the user and the users in the neighborhood, but also considers the similarity between the scenic spots and the scenic spots in the neighborhood, and obtains a neighborhood containing the user. The local neighborhood structure information of the domain and the neighborhood of the scenic spot can enhance the personalization and accuracy of the travel recommendation;
3、本发明在预测用户对景点的得分时，将隐因子模型的全局结构信息与记忆网络的局部邻域结构信息进行结合，充分考虑了全局与局部的信息，从而确保了旅游推荐的有效性。3. The present invention combines the global structure information of the latent factor model with the local neighborhood structure information of the memory network when predicting the user's score for the scenic spot, and fully considers the global and local information, thereby ensuring the validity of the travel recommendation. .
具体实施方式Detailed ways
为使本发明的目的、技术方案和优点更加清楚明白，以下结合具体实例，对本发明进一步详细说明。In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific examples.
一种基于协同记忆网络的旅游推荐方法，具体包括以下步骤：A travel recommendation method based on collaborative memory network, which specifically includes the following steps:
步骤1、利用爬虫工具从旅游网站上采集用户的游记，从中抽取出各个用户所访问过的景点，并对所有用户及其所有景点进行分别编号。Step 1. Use the crawler tool to collect the travel notes of users from the travel website, extract the scenic spots visited by each user, and number all the users and all the scenic spots respectively.
利用现有的网络爬虫手段，在携程、马蜂窝等旅游网站上采集用户的游记，根据用户的游记中所包含的用户去过景点序列的半结构化数据，抽取出各个用户去过的景点，并对采集到的所有用户及其所有景点进行分别编号，达到数据预处理的目的。Using the existing web crawler means, collect users' travel notes on travel websites such as Ctrip and Mafengwo, and extract the scenic spots visited by each user according to the semi-structured data of the sequence of scenic spots the user has visited in the travel notes. All the collected users and all their scenic spots are numbered separately to achieve the purpose of data preprocessing.
对于游记中用户所访问的景点，可以将游记中所有提及的景点全部进行保留，也可以根据用户评分对景点进行适当剔除，即将该用户访问过景点中评分较高的景点保留，例如，当所抽取的景点综合评分是2.8，3.5，4.5等，即五分制时，则将大于等于3分的景点保留下来。For the scenic spots visited by users in the travel notes, all the scenic spots mentioned in the travel notes can be retained, or the scenic spots can be appropriately eliminated according to the user's rating, that is, the scenic spots with higher ratings that the user has visited are reserved. The comprehensive scores of the extracted scenic spots are 2.8, 3.5, 4.5, etc., that is, when the five-point system is used, the scenic spots with a score of 3 or more will be retained.
由于原始的游记数据无法直接用于后续计算，需要为采集到的所有用户和景点进行统一编号，用唯一的ID值来表示，这样不仅便于后续进一步处理，也能对用户信息进行匿名化处理。例如：第一个用户“03202001”的ID设置为0，第二个用户“03202002”的ID设置为1；第一个景点“七星景区”的ID值设置为0，第二个景点“象山景区”的ID设置为1；以此类推，后续数据的ID值逐个加1。Since the original travel data cannot be directly used for subsequent calculations, it is necessary to uniformly number all the collected users and scenic spots, and represent them with a unique ID value, which not only facilitates subsequent further processing, but also anonymizes user information. For example: the ID of the first user "03202001" is set to 0, the ID of the second user "03202002" is set to 1; the ID value of the first scenic spot "Seven Star Scenic Spot" is set to 0, the second scenic spot "Xiangshan Scenic Spot" ” ID is set to 1; and so on, the ID value of subsequent data is incremented by 1 one by one.
步骤2、将每个用户访问过的所有景点都分别作为一个集合，构建关于用户的景点邻域集。将访问过每个景点的所有用户都分别作为一个集合，构建关于景点的用户邻域集。Step 2. Take all the scenic spots visited by each user as a set respectively, and construct a scenic spot neighborhood set about the user. All users who have visited each scenic spot are regarded as a set respectively, and a user neighborhood set about the scenic spot is constructed.
将同一个用户访问过的所有景点通过集合<P1，P2,…Pn>的形式来表示，构成关于该用户的景点邻域集，其中P1,P2,…Pn表示同一个用户访问过的各景点的ID。例如：用户U1游览过景点P1、P2、……Pn，则关于用户U1的景点邻域集为<P1，P2,…Pn>。All scenic spots visited by the same user are represented in the form of a set <P1, P2,...Pn> to form a neighborhood set of scenic spots about the user, where P1, P2,...Pn represent the scenic spots visited by the same user id. For example, if the user U1 has visited the scenic spots P1, P2, ... Pn, the scenic spot neighborhood set about the user U1 is <P1, P2, ... Pn>.
将访问过同一个景点的所有用户通过集合<U1,U2,…Un>的形式来表示，构成关于该景点的用户邻域集，其中U1,U2,…Un表示访问过同一个景点的各用户的ID。例如：用户U1、U2、……Un都游览过景点P1，则关于景点P1的用户邻域集为<U1,U2,…Un>。All users who have visited the same scenic spot are represented in the form of a set <U1, U2,…Un> to form a user neighborhood set about the scenic spot, where U1, U2,…Un represent the users who have visited the same scenic spot id. For example, if the users U1, U2, ...Un have all visited the scenic spot P1, the user neighborhood set about the scenic spot P1 is <U1, U2, ...Un>.
步骤3、利用神经网络将所有用户和景点分别映射到两个特征向量空间中，其中一个特征向量空间包括用户向量矩阵和景点向量矩阵，另一个特征向量空间包括用户外部记忆矩阵和景点外部记忆矩阵。其中用户向量矩阵和用户外部记忆矩阵的维度为n×d，景点向量矩阵和景点外部记忆矩阵的维度为m×d，n，m分别表示用户和景点的个数，d表示给定的矩阵的维度，在本实施例中，d＝50。Step 3. Use neural network to map all users and attractions into two feature vector spaces, one of which includes user vector matrix and attraction vector matrix, and the other feature vector space includes user external memory matrix and attraction external memory matrix. . The dimension of user vector matrix and user external memory matrix is n×d, and the dimension of scenic spot vector matrix and scenic spot external memory matrix is m×d, n, m represent the number of users and scenic spots respectively, d represents the given matrix. dimension, in this embodiment, d=50.
将用户的数量n和景点的数量m这两个参数输入到一个神经网络模块中，神经网络就会输出一个维度为n×d的有关用户的向量矩阵和维度为m×d的有关景点的向量矩阵。由于这2个向量矩阵中的每一个值是通过自己设定的初始化方法产生的(比如正态分布，随机产生一个0～1的数，这个数是随机的)，所以两次将参数n和m输入到神经网络中会得到不同的结果，第一次得到“n×d的用户向量矩阵和m×d的景点向量矩阵”，第二次得到“n×d的用户外部记忆矩阵和m×d的景点外部记忆矩阵”。Input the two parameters of the number of users n and the number of attractions m into a neural network module, the neural network will output a vector matrix of related users with dimension n×d and a vector of related attractions with dimension m×d matrix. Since each value in these two vector matrices is generated by the initialization method set by itself (such as normal distribution, a number between 0 and 1 is randomly generated, this number is random), so the parameters n and If m is input into the neural network, different results will be obtained. The first time you get "n×d user vector matrix and m×d attraction vector matrix", the second time you get "n×d user external memory matrix and m×d d's attractions external memory matrix".
此外，为了使得用户的向量与该用户访问过的景点的向量之间的相似度比与他没访问过的景点向量之间的相似度更高，其中相似度是通过点积的方式来衡量的，在本实施例中，还需要对上述所得到的用户向量矩阵和景点向量矩阵使用广义矩阵分解方法(一种深度学习模型)来进行预训练更新。In addition, in order to make the similarity between the user's vector and the vector of attractions that the user has visited is higher than the similarity between the vector of attractions he has not visited, the similarity is measured by dot product. , in this embodiment, it is also necessary to use a generalized matrix decomposition method (a deep learning model) to perform pre-training update on the obtained user vector matrix and sights vector matrix.
步骤4、计算用户向量矩阵中的用户向量与其用户邻域集中其他用户向量的相似度，作为权重系数，然后根据用户向量矩阵、相应的权重系数以及用户外部记忆矩阵构建用户记忆网络，获得关于用户的邻域结构信息。Step 4. Calculate the similarity between the user vector in the user vector matrix and other user vectors in the user neighborhood set as a weight coefficient, and then construct a user memory network according to the user vector matrix, the corresponding weight coefficient and the user external memory matrix, and obtain information about the user. the neighborhood structure information.
在统一编号的用户集中选择一个编号为u的用户，在统一编号的景点集中选择一个编号为i的景点，将用户u和景点i作为当前记忆网络构建的用户和景点。Select a user numbered u in the set of uniformly numbered users, select a scenic spot numbered i in the set of uniformly numbered scenic spots, and take user u and scenic spot i as the users and scenic spots constructed by the current memory network.
首先，在用户向量矩阵中，将用户u所对应的向量与其关于景点i的用户邻域集中的各个用户向量进行点乘，再将点乘的结果加上景点i所对应的景点向量矩阵中的向量与该用户邻域集中的用户向量的点乘，得到用户u与其用户邻域集中各个用户的相似度quiv：First, in the user vector matrix, the vector corresponding to user u is multiplied by each user vector in the user neighborhood set about the scenic spot i, and then the result of the dot multiplication is added to the scenic spot vector matrix corresponding to the scenic spot i. The dot product of the vector and the user vector in the user neighborhood set can obtain the similarity q uiv between the user u and each user in the user neighborhood set:
式中，quiv表示关于景点i的用户邻域集中用户u和用户v在归一化前的相似度，mu，mv分别表示用户向量矩阵中用户u和用户v所对应的特征向量，ei表示景点向量矩阵中景点i所对应的特征向量，N(i)表示所有访问过景点i的用户所构成的用户邻域集；In the formula, quiv represents the similarity between user u and user v in the user neighborhood set of scenic spot i before normalization, m u , m v represent the feature vectors corresponding to user u and user v in the user vector matrix, respectively, e i represents the feature vector corresponding to the scenic spot i in the scenic spot vector matrix, and N(i) represents the user neighborhood set composed of all users who have visited the scenic spot i;
接着，将上述所得到的用户u与其用户邻域集中各个用户的相似度quiv进行归一化，并将归一化后的相似度作为用户u的权重系数puiv：Next, normalize the similarity q uiv between the user u and each user in the user neighborhood set obtained above, and use the normalized similarity as the weight coefficient p uiv of the user u:
式中，exp(x)是以e为底x为指数的指数函数。In the formula, exp(x) is an exponential function with e as the base and x as the exponent.
最后，将用户u的权重系数puiv作为注意力机制中的权重，并将关于景点i的用户邻域集中各个用户v在用户外部记忆矩阵中的特征向量进行加权求和，通过用户外部记忆矩阵中用户u的用户邻域集中的用户向量来表示用户u，以此来获得包含用户邻域信息的向量表示oui，即关于用户u的邻域结构信息。其中：Finally, the weight coefficient p uiv of user u is used as the weight in the attention mechanism, and the eigenvectors of each user v in the user's external memory matrix in the user's neighborhood set about the scenic spot i are weighted and summed, and through the user's external memory matrix User u is represented by the user vector in the user neighborhood set of user u in , so as to obtain the vector representation o ui containing user neighborhood information, that is, the neighborhood structure information about user u. in:
式中，puiv表示用户u的权重系数，cv表示关于景点i的用户邻域集中各个用户v在用户外部记忆矩阵中的特征向量。In the formula, p uiv represents the weight coefficient of user u, and cv represents the feature vector of each user v in the user's external memory matrix in the user neighborhood set about the scenic spot i.
步骤5、计算景点向量矩阵中的景点向量与其景点邻域集中的其他景点向量的相似度，作为权重系数，然后根据景点向量矩阵、相应的权重系数以及景点外部记忆矩阵构建景点记忆网络，获得关于景点的邻域结构信息。Step 5. Calculate the similarity between the scenic spot vector in the scenic spot vector matrix and other scenic spot vectors in the scenic spot neighborhood set as the weight coefficient, and then construct the scenic spot memory network according to the scenic spot vector matrix, the corresponding weight coefficient and the scenic spot external memory matrix, and obtain the information about the scenic spot. Information about the neighborhood structure of attractions.
在统一编号的用户集中选择一个编号为u的用户，在统一编号的景点集中选择一个编号为i的景点，将用户u和景点i作为当前记忆网络构建的用户和景点。Select a user numbered u in the uniformly numbered user set, select a scenic spot numbered i in the uniformly numbered scenic spot set, and take user u and scenic spot i as the users and scenic spots constructed by the current memory network.
首先，在景点向量矩阵中，将景点i所对应的向量与其关于用户u的景点邻域集中的各个景点向量进行点乘，再将点乘的结果加上用户u所对应的用户向量矩阵中的向量与该景点邻域集中的景点向量的点乘，得到景点i与其景点邻域集中各个景点的相似度qiuj：First, in the scenic spot vector matrix, the vector corresponding to scenic spot i is multiplied by each scenic spot vector in the scenic spot neighborhood set about user u, and then the result of the dot product is added to the user vector matrix corresponding to user u. The dot product of the vector and the scenic spot vector in the scenic spot neighborhood set can get the similarity q iuj between the scenic spot i and each scenic spot in the scenic spot neighborhood set:
式中，qiuj表示关于用户u的景点邻域集中景点i和景点j在归一化前的相似度，ei，ej分别表示景点向量矩阵中景点i和景点j所对应的特征向量，mu表示用户向量矩阵中用户u所对应的特征向量，S(u)表示用户u访问过的所有景点所构成的景点邻域集；In the formula, q iuj represents the similarity between the scenic spot i and the scenic spot j in the scenic spot neighborhood set of user u before normalization, e i , e j represent the feature vector corresponding to the scenic spot i and the scenic spot j in the scenic spot vector matrix, respectively, m u represents the feature vector corresponding to user u in the user vector matrix, and S(u) represents the scenic spot neighborhood set composed of all scenic spots visited by user u;
接着，将景点i与其景点邻域集中各个景点的相似度qiuj进行归一化，归一化后的相似度作为景点i的权重系数piuj：Next, normalize the similarity qiuj between the scenic spot i and each scenic spot in its neighborhood set, and use the normalized similarity as the weight coefficient p iuj of the scenic spot i:
式中，exp(x)是以e为底x为指数的指数函数。In the formula, exp(x) is an exponential function with e as the base and x as the exponent.
最后，将景点i的权重系数piuj作为注意力机制中的权重，并将用户u的景点领域集中各个景点j在景点外部记忆矩阵中的特征向量进行加权求和，即通过景点外部记忆矩阵中景点i的景点邻域集中的景点向量来表示景点i，以此来获得包含景点邻域信息的向量表示oiu，即关于景点i的邻域结构信息。其中：Finally, the weight coefficient p iuj of the scenic spot i is used as the weight in the attention mechanism, and the eigenvectors of each scenic spot j in the scenic spot j in the scenic spot field of user u are weighted and summed, that is, through the external memory matrix of the scenic spot The scenic spot vector in the scenic spot neighborhood set of the scenic spot i is used to represent the scenic spot i, so as to obtain the vector representation o iu containing the scenic spot neighborhood information, that is, the neighborhood structure information about the scenic spot i. in:
式中，piuj表示景点i的权重系数，yj表示关于用户u的景点邻域集中各个景点j在景点外部记忆矩阵中的特征向量。In the formula, p iuj represents the weight coefficient of the scenic spot i, and y j represents the feature vector of each scenic spot j in the external memory matrix of the scenic spot in the scenic spot neighborhood set of user u.
步骤6、将用户向量矩阵中用户u所对应的特征向量与景点向量矩阵中景点i所对应的特征向量进行元素积，得到用户对景点的偏好信息的向量表示，即隐因子模型的用户u对景点i的全局结构信息；Step 6. Perform element product of the feature vector corresponding to user u in the user vector matrix and the feature vector corresponding to the scenic spot i in the scenic spot vector matrix to obtain the vector representation of the user's preference information for the scenic spot, that is, the user u pair of the hidden factor model. The global structure information of the scenic spot i;
步骤7、通过非线性的方式将隐因子模型的全局结构信息与包含用户邻域信息的向量表示(关于用户u的邻域结构信息oui)和包含景点邻域信息的向量表示(关于景点i的邻域结构信息oiu)进行拼接。Step 7. In a nonlinear way, combine the global structure information of the latent factor model with the vector representation containing the user neighborhood information (about the neighborhood structure information o ui of the user u ) and the vector representation containing the scenic spot neighborhood information (about the scenic spot i ). The neighborhood structure information o iu ) is spliced.
例如，隐因子模型的全局结构信息、包含用户邻域信息的向量表示及包含景点邻域信息的向量表示的向量维度都是100维，则拼接后的向量维度为300维。For example, the vector dimension of the global structure information of the latent factor model, the vector representation containing the user neighborhood information, and the vector representation containing the scenic spot neighborhood information are all 100 dimensions, then the spliced vector dimension is 300 dimensions.
步骤8、将拼接后的向量输入到多层神经网络中去得到用户对景点的最终评分，评分值越小，说明用户去游览该景点的可能性越小，评分值越大，说明用户越有可能游览该景点。具体公式如下：Step 8. Input the spliced vector into the multi-layer neural network to obtain the user's final rating for the scenic spot. The smaller the rating value, the less likely the user is to visit the scenic spot, and the larger the rating value, the more likely the user is. Possibility to visit this attraction. The specific formula is as follows:
rui＝vTφ(U(mu⊙ei)+Woui+Woiu+b) (7)r ui =v T φ(U(m u ⊙e i )+Wo ui +Wo iu +b) (7)
其中，rui表示用户对景点的评分，U、W、v以及b都是神经网络中需要学习的参数，⊙表示元素积，即两个向量的对应元素相乘，得到与原向量相同维度的新的向量，φ(x)表示非线性激活函数。Among them, r ui represents the user's rating of the scenic spot, U, W, v and b are all parameters that need to be learned in the neural network, ⊙ represents the element product, that is, the corresponding elements of the two vectors are multiplied to obtain the same dimension as the original vector. The new vector, φ(x), represents the nonlinear activation function.
步骤9、通过上述一系列方法计算出目标用户对所有景点的评分值，将得到的评分进行排序，选出排名靠前的K个景点，即得到关于该目标用户的Top-K旅游景点推荐。Step 9: Calculate the score values of the target user for all scenic spots through the above series of methods, sort the obtained scores, and select the top K scenic spots to obtain the Top-K tourist attractions recommendation for the target user.
本发明将基于隐因子模型的协同过滤方法与基于邻域或记忆的协同过滤方法进行了融合，能够利用隐因子模型的全局结构信息以及邻域或记忆的局部结构信息的优势，以此达到个性化旅游推荐的目的。The invention integrates the collaborative filtering method based on the latent factor model and the collaborative filtering method based on the neighborhood or memory, and can take advantage of the global structure information of the latent factor model and the local structure information of the neighborhood or memory, so as to achieve individuality. Purpose of travel recommendation.
需要说明的是，尽管以上本发明所述的实施例是说明性的，但这并非是对本发明的限制，因此本发明并不局限于上述具体实施方式中。在不脱离本发明原理的情况下，凡是本领域技术人员在本发明的启示下获得的其它实施方式，均视为在本发明的保护之内。It should be noted that, although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, so the present invention is not limited to the above-mentioned specific embodiments. Without departing from the principles of the present invention, all other embodiments obtained by those skilled in the art under the inspiration of the present invention are deemed to be within the protection of the present invention.
Claims (5)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010618323.2A CN111882381B (en) | 2020-06-30 | 2020-06-30 | A travel recommendation method based on collaborative memory network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010618323.2A CN111882381B (en) | 2020-06-30 | 2020-06-30 | A travel recommendation method based on collaborative memory network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111882381A true CN111882381A (en) | 2020-11-03 |
| CN111882381B CN111882381B (en) | 2022-06-03 |
Family
ID=73156626
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010618323.2A Active CN111882381B (en) | 2020-06-30 | 2020-06-30 | A travel recommendation method based on collaborative memory network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111882381B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112712418A (en) * | 2021-03-25 | 2021-04-27 | 腾讯科技（深圳）有限公司 | Method and device for determining recommended commodity information, storage medium and electronic equipment |
| CN112749345A (en) * | 2021-02-09 | 2021-05-04 | 上海海事大学 | K nearest neighbor matrix decomposition recommendation method based on neural network |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180103052A1 (en) * | 2016-10-11 | 2018-04-12 | Battelle Memorial Institute | System and methods for automated detection, reasoning and recommendations for resilient cyber systems |
| US20190073363A1 (en) * | 2017-09-07 | 2019-03-07 | Xerox Corporation | Contextual memory bandit for proactive dialogs |
| CN110008408A (en) * | 2019-04-12 | 2019-07-12 | 山东大学 | A session recommendation method, system, device and medium |
| CN110263256A (en) * | 2019-06-21 | 2019-09-20 | 西安电子科技大学 | Personalized recommendation method based on multi-modal heterogeneous information |
| CN110413844A (en) * | 2019-05-24 | 2019-11-05 | 浙江工业大学 | Dynamic link prediction method based on spatiotemporal attention depth model |
| CN110825884A (en) * | 2019-11-05 | 2020-02-21 | 腾讯科技（深圳）有限公司 | Embedded representation processing method and device based on artificial intelligence and electronic equipment |
| CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Implementation and Classification of Bilinear Graph Neural Network Models for Modeling Neighbor Interactions |
| CN111061856A (en) * | 2019-06-06 | 2020-04-24 | 北京理工大学 | Knowledge perception-based news recommendation method |
| CN111222054A (en) * | 2020-01-03 | 2020-06-02 | 中国计量大学 | A Conversational Social Recommendation Method Based on Contextual Neighbor Relationship Modeling |
-
2020
- 2020-06-30 CN CN202010618323.2A patent/CN111882381B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180103052A1 (en) * | 2016-10-11 | 2018-04-12 | Battelle Memorial Institute | System and methods for automated detection, reasoning and recommendations for resilient cyber systems |
| US20190073363A1 (en) * | 2017-09-07 | 2019-03-07 | Xerox Corporation | Contextual memory bandit for proactive dialogs |
| CN110008408A (en) * | 2019-04-12 | 2019-07-12 | 山东大学 | A session recommendation method, system, device and medium |
| CN110413844A (en) * | 2019-05-24 | 2019-11-05 | 浙江工业大学 | Dynamic link prediction method based on spatiotemporal attention depth model |
| CN111061856A (en) * | 2019-06-06 | 2020-04-24 | 北京理工大学 | Knowledge perception-based news recommendation method |
| CN110263256A (en) * | 2019-06-21 | 2019-09-20 | 西安电子科技大学 | Personalized recommendation method based on multi-modal heterogeneous information |
| CN110825884A (en) * | 2019-11-05 | 2020-02-21 | 腾讯科技（深圳）有限公司 | Embedded representation processing method and device based on artificial intelligence and electronic equipment |
| CN110826700A (en) * | 2019-11-13 | 2020-02-21 | 中国科学技术大学 | Implementation and Classification of Bilinear Graph Neural Network Models for Modeling Neighbor Interactions |
| CN111222054A (en) * | 2020-01-03 | 2020-06-02 | 中国计量大学 | A Conversational Social Recommendation Method Based on Contextual Neighbor Relationship Modeling |
Non-Patent Citations (1)
| Title |
|---|
| 陈细玉等: "一种融合注意力和记忆网络的序列推荐算法", 《信息技术与网络安全》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112749345A (en) * | 2021-02-09 | 2021-05-04 | 上海海事大学 | K nearest neighbor matrix decomposition recommendation method based on neural network |
| CN112749345B (en) * | 2021-02-09 | 2023-11-14 | 上海海事大学 | A K-nearest neighbor matrix decomposition recommendation method based on neural network |
| CN112712418A (en) * | 2021-03-25 | 2021-04-27 | 腾讯科技（深圳）有限公司 | Method and device for determining recommended commodity information, storage medium and electronic equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111882381B (en) | 2022-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111428147B (en) | A social recommendation method with heterogeneous graph convolutional networks combining social and interest information | |
| CN108846422B (en) | Account number association method and system across social networks | |
| CN110287336B (en) | Tourist map construction method for tourist attraction recommendation | |
| CN107133277B (en) | A Tourist Attraction Recommendation Method Based on Dynamic Topic Model and Matrix Factorization | |
| CN107133262B (en) | A Personalized POI Recommendation Method Based on Multi-influence Embedding | |
| CN105069717A (en) | Personalized travel route recommendation method based on tourist trust | |
| CN110097125A (en) | A kind of across a network account correlating method indicated based on insertion | |
| CN113505204B (en) | Recall model training method, search recall device and computer equipment | |
| CN109743196B (en) | Network characterization method based on cross-double-layer network random walk | |
| CN109992674B (en) | Recommendation method fusing automatic encoder and knowledge graph semantic information | |
| CN113486190A (en) | Multi-mode knowledge representation method integrating entity image information and entity category information | |
| CN108304429A (en) | Information recommendation method, device and computer equipment | |
| CN107577782B (en) | Figure similarity depicting method based on heterogeneous data | |
| CN110119475A (en) | A kind of POI recommended method and recommender system | |
| CN115408621B (en) | Point-of-interest recommendation method considering linear and nonlinear interaction of auxiliary information features | |
| CN111882381B (en) | A travel recommendation method based on collaborative memory network | |
| CN113095948A (en) | Multi-source heterogeneous network user alignment method based on graph neural network | |
| CN113779429A (en) | Traffic congestion situation prediction method, device, equipment and storage medium | |
| CN112559658A (en) | Address matching method and device | |
| CN112685541A (en) | Social media rumor detection method based on multitask learning | |
| Zhang et al. | FM-based: algorithm research on rural tourism recommendation combining seasonal and distribution features | |
| CN110532464A (en) | A kind of tourism recommended method based on more tourism context modelings | |
| CN116823535A (en) | Journey planning and intelligent navigation system based on multi-mode large model | |
| CN112800111A (en) | A Location Prediction Method Based on Training Data Mining | |
| CN109960755A (en) | A User Privacy Protection Method Based on Dynamic Iterative Fast Gradient |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |