本站原创
摘要： 为了解决电子商务在线营销中的信息阻塞等理由，提出了一种新型的、高效的推荐系统模型。给出了该系统模型中采用的疯狂蚂蚁算法，并描述了它的消息结构和处理流程，最后给出了一个系统应用实例和性能对比结果。仿真实验表明该模型在实时性和客户满意度上超过了传统系统。
关键词： 电子商务；推荐模型；蚁群算法；匹配；算法
1006-4311（2013）32-0014-03
0 引言
在过去十年中，电子商务已经成为了一个集成的系统工程；其中海量的营销网页使得客户的选择变得低效而无趣——客户过去面对的商品信息较为有限，而今在网上信息爆炸导致了客户的信息过载理由。为解决这一理由，互联网研究人员采用了诸多办法提高客户的信息检索效率；近年来，网店也应用了大量的有效策略改善其营销策略，其中一种策略就是获取和分析客户的偏好信息，从而主动向客户提供他们感兴趣的商品目录等信息，而不是被动等待客户自己检索相关信息；经过一定时间的实验与应用，这种被称之为“推荐”的在线营销策略，已经证明是信息过载的一种有效解决手段。
总的来说，推荐系统是根据客户的个性化需求、兴趣习惯等信息，将客户感兴趣的商品、怎么写作等相关信息推荐给客户的个性化信息智能处理系统。和搜索引擎相比，它能够通过研究客户的兴趣偏好，进行个性化处理，由系统发现客户的兴趣点，从而引导其发现自己的信息需求。目前这些系统已经广泛应用于电影、音乐、网页等方面的在线营销。
进来，越来越多的推荐系统尝试使用各类先进算法提高其推荐效果；根据基础算法的不同，可将推荐系统分为三类：基于内容的推荐、协同过滤推荐以及组合推荐。目前，协同过滤推荐系统最为流行。协同过滤推荐系统往往采用最近邻匹配算法，利用客户的历史喜好信CACER：一种基于疯狂蚁群算法的电子商务推荐模型由提供海量免费论文范文的www.udooo.com,希望对您的论文写作有帮助.息计算客户之间的距离，然后利用目标客户的最近邻居客户对商品评价的加权评价的绝对值来判断客户对特定商品或怎么写作的喜好程度，从而根据这一喜好程度来对目标客户进行推荐。协同过滤对推荐对象没有特殊的要求，并且能处理非结构化的复杂对象，如音乐、电影，因此在电子商务领域应用较广。
蚁群优化算法（ACO）是一个相对较新的启发式算法，能够有效的进行局部搜索，并解决相应组合优化理由，如推荐等[8]；该算法模拟了自然界中蚁群寻找食物的过程，从而寻找理由的最有效的途径。第一个蚁群算法是由clolrni等人开发，并成功地应用于旅行商理由（TSP）。而bulleneimer等人基于这种基本思路，设计了路径选择优化解决方案，并随之提出了一种改善的蚁群算法。因为推荐理由是非常复杂的，而基本的蚁群算法不能直接应用，并满意的系统的性能需要；因此，许多研究人员都提出了新的策略来提高改善方案来解决一系列不同的理由。
上述研究成果，成功证明了蚁群算法与其他算法来解决各类大规模复杂理由的可能性；基于他们的启发，本文提出了一种新型的电子商务推荐系统模型（CACER： crazy ant colony E-commerce recommendation）。本文其他部分安排如下：第二部分总结传统推荐系统的不足，并提出系统模型；第三部分，给出疯狂蚁群算法及其应用，包括算法数据结构、运转流程等。最后两部分分别给出仿真实验结果，并总结全文。
①基于内容的推荐系统：此类系统能够根据客户对内容的偏好，自动检索过滤众多内容，选择一些与客户偏好匹配度高的信息推荐给客户，并能够依据相似度的高低顺序，产生推荐商品列表。由于其信息依赖程度较低，因此被应用在NewWeeder、Tapestry、InfoFinder、Mooney等系统中。
②基于协同过滤的推荐系统：此类系统基于一组兴趣相同的客户或项目进行推荐，它根据邻居客户（与目标客户兴趣接近或相似的客户）的偏好信息产生对目标客户的推荐列表，因此适合于客户数量较大的电子商务系统，目前已被应用在Amazon、Linden、Smith、and York、Sarwar： Sarwar、Karypis、Konstan、Riedl、Hofmann等系统中。
③组合推荐系统按不同的需求，采取不同的混合策略将不同推荐类型或推荐算法进行组合并推荐。例如，Adomicius和Tuzhilin等提出了一种基于上下文感知的推荐系统，该系统将传统的“用户-项目”二维评分效用模型拓展为包含多种上下文信息的多维评分效用模型，因此取得了较好的精度和客户满意度。
我们的系统模型基于内容推荐，并根据疯狂蚁群算法进行了改造和创新，系统的整体结构描述如下（图1）。该模型中，应用了部分的传统模块用于实现推荐系统，例如商品的联机事务处理（OLTP）、数据挖掘、商务统计与在线数据库等。整个系统的核心是基于疯狂蚁群算法的内容推荐。
①该系统运营的第一阶段是在客户偏好分析模块中运转的，即：在商品的浏览、订货、购写等过程中的将客户群体细分成不同的集群的推荐处理模块；相关信息保存在电子商务系统的基本信息数据库（交易数据等内容）。
②客户产生的实时信息，包括：客户在线检索和商品搜索选项等，将分类并保存到在线数据库，为后续工作提供交易信息等信息支持。
③B/C OLTP（联机事务处理）和数据挖掘模块集成了电子商务的基本信息和实时交易信息推荐的所有内容，包括客户及相关商品的信息。
④推荐处理模块，该模块是系统的核心，它通过上述模块采集并预处理过后的集成数据，生成推荐列表；疯狂蚁群算法是其中的关键。该算法搜索客户与商品的匹配指数从两个方向上生成推荐列表。首先是根据客户的偏好生成推荐列表；其次是当新商品上架，也可根据客户的需求信息，向客户进行推荐，提供自动选择倡议怎么写作。⑤推荐通信模块，该模块的推荐实施策略包括电子邮件、在线广告、弹出式通知等等，这也是一个双向通信模块，即：它不但可以推荐的商品列表给客户，而且可以跟踪生成新商品的特征词，当客户对新商品进行评价或购写新商品时，该模块可以通过智能推理算法预测的该商品归于于哪一类客户的偏好。
上述模块与本模型，均能够通过较小的修改，嵌入其他推荐系统。而疯狂蚁群推荐算法也可以用于替换其他基于内容的推荐算法，具体细节参见下文。
我们的算法可以利用3种蚂蚁Agent处理推荐理由。
第一种是侦察（Scout）蚂蚁，它被派出去收集新推荐商品吸引程度。第二种是承载（Carrier）蚂蚁，它被派出找到客户，并产生推荐商品列表，或者根据新商品，产生推荐客户列表。最后是推销员（Sale）蚂蚁，它主要用于生成随机推荐广告。各种蚂蚁模型描述如下图2所示。
随着客户历史信息的不断丰富，侦察蚂蚁可以从B/C OLTP和数据发掘模块开始着手，进入电子商务基本数据库，通过交易统计模块等提供预处理怎么写作。在他们的搜索过程中，收集和分类新客户的信息。进入电子商务基本数据库后，就可以使用客户购物兴趣等信息素，在各类商品的信息项上记录关键词和描述词（信息素）。最终，这些蚂蚁将回到B/C OLTP和数据挖掘模块，汇总“爬行”过的商品的客户吸引程度；数据发掘模块将由此发现商品和客户之间的关系，及其变CACER：一种基于疯狂蚁群算法的电子商务推荐模型由优秀论文网站www.udooo.com提供,助您写好论文.化趋势；并生成初级的商品推荐目录。
当注册客户进入电子商务系统，承载蚂蚁从推荐处理模块（蚁巢）出发，为注册客户提供全程跟踪怎么写作。它们不但可以收集和发现客户的个性化的需求信息。偏好等信息将被它们收集起来，与商品信息中携带的信息素进行匹配，并选择信息素浓度较高的商品进行更新，并注入电子商务基础数据库。最终，承载蚂蚁的信息浓度达到一定程度后，退回推荐处理模块，根据采集到的信息，向客户推荐相关商品。
当未注册的用户进入系统和浏览上平项目时，推销员（Sale）蚂蚁从推荐处理模块（蚁巢）出发，而未注册客户的的个性化信息均由其收集后进行分类。此外，此类蚂蚁可以收集和发现未注册客户的兴趣点、客户的浏览序列。收集到的偏好等信息，将用于相关信息素的更新；当这些推销员蚂蚁结束处理，进入了电子商务基本数据库，汇总并更新相关商品信息后；这些信息将通过B/C OLTP和数据挖掘模块，优化推荐列表的生成过程和结果。
在这个实例中，一只蚂蚁带着若干客户兴趣信息进入了该二维空间，它将这些信息素标识在空间中。其他蚂蚁随后陆续到达该空间，对若干商品进行了信息素标识。当两类信息素被标识完毕后，一只销售蚂蚁或者承载蚂蚁从中收集了所需的TopN商品列表，然后将其交给推荐实施模块或直接推荐给未注册客户。
In this paper， we present a novel E-commerce recommendation model called CACER （crazy ant colony E-commerce recommendation）. The rest of the paper is organized as follows. Partition 2 makes a summary of traditional models and gives a distributed E-commerce recommendation model. In partition 3， a crazy ant colony algorithm is given and its application is proposed. And in the section， the message and data structuresCACER：一种基于疯狂蚁群算法的电子商务推荐模型由优秀论文网站www.udooo.com提供,助您写好论文. of CACER are given. Then its work flow is described in detail. In partition 4， correctness and efficiency of the algorithm are proved. And compared with a traditional， simulation results will be given in partition 5. At last， partition 6 concludes the paper.
II. SYSTEM MODEL
Based on how recommendations are made， recommender systems are usually classified into the following three categories[8-11]：
Content-based recommendations： Recommendations are provided by automatically matching a customer's interests with items' contents. Items that are similar to ones the user preferred in the past are now recommended. Notice that recommendations are made without relying on information provided by other customers， but solely on items’ contents and users’ profiles. NewsWeeder applied this method to build a net-news filtering system. Other applications include Tapestry， InfoFinder， Mooney， and so on.Collaborative filtering： Recommendations are made for items that people with similar tastes and preferences liked in the past. This technique is widely used and is the preferred method for personal recommendation. Many systems， such as Amazon.com： Linden， Smith， and York， Sarwar： Sarwar， Karypis， Konstan， and Riedl， Hofmann， and they he adopted this technique.
Hybrid approaches： These approaches combine collaborative and content-based methods. Fab is a hybrid contentbased， collaborative webpage recommendation system that eliminates many handicaps of the pure versions of either approach. According to the classification scheme proposed by Adomicius & Tuzhilin， hybrid recommendation systems can be divided into four categories： （1） implementing collaborative and content-based methods separately and combining their predictions， （2） incorporating some content-based characteristics into a collaborative approach， （3） incorporating some collaborative characteristics into a content-based approach， and （4） constructing a general， unifying model that incorporates both collaborative and content-based characteristics.
Then our system model is based on Content-based recommendation with crazy ant algorithms. And it can be figured with Fig.1.
In the model， some traditional modules are utilized to implement recommendation systems， such as Product OLTP and Data mine， Commerce Statistic and On-line DB. And the recommendation processing module is the core of the system model with crazy ant colony algorithms.
In the addition， the system work flow can be described as following.
（1） The first phase is a customer interest profiling step， where product bought by a customer in the past is segmented into different clusters by recommendation processing modules. And related information is committed into these sub-DBs of E-commerce basic information DB， including transaction data and so on.
（2） Some real-time information， including customer retrieval and search items， can be classified and sed into the on-line DB， which provides transaction information and so on to the transaction statistics module.
（3） B/C OLTP and Data mine module integrates E-commerce basic information and real-time transaction information into recommendation items， including customer and related product information.
（4） Recommendation processing module， as our system core， integrates all kinds of data. And crazy ant colony algorithms can be executed by it. Furthermore， the algorithm will search customer-product matching metrics. ThCACER：一种基于疯狂蚁群算法的电子商务推荐模型由提供海量免费论文范文的www.udooo.com,希望对您的论文写作有帮助.en the module can provide recommendation lists， including customer items and interested product items. Also according to customer information， it can automatically choose specific methods to recommend. Through E-commerce push communication， related information can be tranerred to the customers.（5） Push communication module， composed of Email， NoCACER：一种基于疯狂蚁群算法的电子商务推荐模型相关论文由www.udooo.com收集,如需论文查抄袭率.te， and so on， can be bidirectional sub-module. Mainly， it can recommend product items to customers. Subsequently， the new product features are presented to it and the preference value that the customer possibly has for the new product can be predicted by the generalization inference ability of intelligent algorithms. And through other modules， the preference prediction values for all new products are ranked and the top-n items with the highest values will be recommended to the customer.
Note that our model can be extended for other recommendation systems with little modification. And also， crazy ant colony algorithm can be utilized to replace other content-based recommendation algorithms in the system models. And the details of them can be described as following sections.
III. CRAZY ANT COLONY ALGORITHM
In the section， the main idea， model， data structures， and implementation models are given for crazy ant colony algorithm.
A. Main Idea and Model
The crazy ant colony algorithm is a particular algorithm of ant colony optimization （ACO） which is based on agents that simulate the natural behior of ants， develop mechanis for cooperation， and assist them in using experience to find the shortest path between a food source and the nest. ACS is a population-based heuristics that enables the exploration of the positive feedback whereas the ants are able to communicate （ants lay pheromone for indirect communication） information concerning food source via an aromatic essence. The ants lay pheromone and heuristic information to mark trails. As the paths are visited by other ants， some of the trails may be reinforced and others paths may be allowed to evaporate. Pheromone trails can be observed via the number of ants passing through the trail. When there are more pheromones on a path， there is larger probability that other ants will use that path， and therefore the pheromone trail on such a path will grow faster and attract more ants to follow （so called positive feedback）. An iterative local search algorithm tries to search the current paths to neighboring paths until a better solution is found. And the main idea of CAC algorithm can be imagined as following：
Our algorithm can utilize three kinds of ant agents to deal with the recommendation problem. The first is ascout ant， which can be dispatched to collect the attracting trends of new recommendation products. The others are a band of carrier ants， which are sent to find recommendation products for customers. And the lasts are some salean ants， which randomly recommend AD products to customer. And all kinds of ant models are described as following and showed in .figure 2.论文资料由论文网www.udooo.com提供,转载请保留地址.The proposed technology flow divides the recommendation process into two phases. The first phase is a customer interest profiling step， where product bought by a customer in the past is standardized and segmented into different clusters in a two-dimensional space. Each resultinCACER：一种基于疯狂蚁群算法的电子商务推荐模型相关范文由写论文的好帮手www.udooo.com提供,转载请保留.g cluster corresponds to an interest area of the customer. In the second phase， the tasks of product preference prediction and recommendation generation are carried out. For a new product that is not purchased by the customer yet， CACER categorizes it into the right cluster first. Then the products included in this cluster， together with their ratings given by the customer explicitly or implicitly， are used the same two-dimensional space. Subsequently， the product attributes are formalized and presented to the two-dimensional space. Then the preference value that the customer possibly has for the product can be predicted by the generalization inference ability of top match algorithm. At last， the preference prediction values for all new products are ranked and the top-n items with the highest values will be recommended to the customer. Figure 3 depicts the processing flows of the technology.
In the example， one ant entered into the two-dimensional space with some customer interests， and it labeled some pheromones in the space. Then other ants arrived in the space with some product attributes as their pheromones. When two kinds of pheromones were labeled in the superposed space， a salean ant or a carrier ant could collect asimilarity top-N product list. Further， it pushed the list to recommendation processing module and recommend the list to customers.
IV. SIMULATION RESULTS
In order to validate the proposed CACE algorithm， a simulation experiment was carried out. The simulation configuration for the experiment was as follows： First， a product dataset was designed by randomly choosing 500 products. The products in this category represent the most popular items that are often bought by customers from Internet. The behior of a typical customer was simulated by randomly assigning 100 products out of the 500 products with ratings between 1 and 10， and the rest with click， se and purchase number. Then doublel experiements he been respectivly executed with CACER and without it.
Figure 4A shows that CACER can make the systme attract more consumers’ attention than before. And in each simulation group， consumers spend more time in AD webs with CACER than without it. The performance of each model is evaluated in terms of recommendation precision， which measures the percentage that products recommended to a customer are actually liked by the customer. And the simulation results show the delays， in which comsumer browse the recommended products.Futhermore， CACER can increase E-commerce turnovers. And the simulation results （in Fig.4 B） show the turnovers he been more about 12% than before. This can be explained by the unique properties exhibited by the CACER model such as user profile with multiple interests， adaptability to customer interest changes and the good robust learning and generalization inference ability exhibited by three kinds of ants.V. CONCLUSION
In this paper， we propose a CACER recommendation system model to help e-commerce websites provide better personalization service for their customers. The experimental results show that the our recommendation model is superior to the before. In addition， the model should be applied to a real e-commerce eCACER：一种基于疯狂蚁群算法的电子商务推荐模型由专注毕业论文与职称论文的www.udooo.com提供,转载请保留.nvironment where customers give their actual ratings to products for validating its practical effect.
REFERENCES：
[1]Deng Zhi-Hong， Wang Zhong-Hui， Zhang Jie. ROBIN： A novel personal recommendation model based on information propagation [J]. EXPERT SYSTEMS WITH APPLICATIONS， 2013，40（13）：5306-5313.
[2]Jing Y， Liu H. Integrating ontology technology with folksonomies for personalized social tag recommendation [J]. APPLIED SOFT COMPUTING，2013，13（8）： 3745-3750.
[3]Zhao Xiangyu， Niu Zhendong， Chen Wei. Interest before liking： Two-step recommendation approaches [J]. KNOWLEDGE-BASED SYSTEMS，2013，48（8）：57-69.
[4]Hernando Antonio， Bobadilla Jesus， Ortega Fernando. Trees for explaining recommendations made through collaborative filtering [J]. INFORMATION SCIENCES，2013，239（8）：1-17.
[5]Lai Chin-Hui， Liu Duen-Ren， Lin Cai-Sin. Novel personal and group-based trust models in collaborative filtering for document recommendation [J]. 2013，239（8）：31-49.
[6]Reimann Jan， Seifert Mirko， Asann Uwe. On the reuse and recommendation of model refactoring specifications [J]. SOFTWARE AND SYSTEMS MODELING，2013，12（3）：579-596.
[7]Lucas Joel P， Luz Nuno Moreno， Maria N. A hybrid recommendation approach for a touri system [J]. EXPERT SYSTEMS WITH APPLICATIONS，2013，40（9）：3532-3550.
[8]Agarwal Deepak， Chen Bee-Chung， Elango Pradheep. Content Recommendation on Web Portals [J]. COMMUNICATIONS OF THE ACM. 2013，56（6）：92-101.
[9]Li Lin， Zhong Luo， Yang Zhenglu. QUBIC： An adaptive approach to query-based recommendation [J]. JOURNAL OF INTELLIGENT INFORMATION SYSTEMS，2013，40（3）：556-587.
[10]Fernandez-Luna Juan M， Huete Juan F， Castells Pablo. Personalization and Recommendation in Information Access [J]. INFORMATION PROCESSING & MANAGEMENT，2013，49（3）：637-639.
[11]Yuan Xiaofang， Lee Ji-Hyun， Kim Sun-Joong. Toward a user-oriented recommendation system for real estate websites[J]. INFORMATION SYSTEMS，2013，38（2）：231-243.
关键词： 电子商务；推荐模型；蚁群算法；匹配；算法
1006-4311（2013）32-0014-03
0 引言
在过去十年中，电子商务已经成为了一个集成的系统工程；其中海量的营销网页使得客户的选择变得低效而无趣——客户过去面对的商品信息较为有限，而今在网上信息爆炸导致了客户的信息过载理由。为解决这一理由，互联网研究人员采用了诸多办法提高客户的信息检索效率；近年来，网店也应用了大量的有效策略改善其营销策略，其中一种策略就是获取和分析客户的偏好信息，从而主动向客户提供他们感兴趣的商品目录等信息，而不是被动等待客户自己检索相关信息；经过一定时间的实验与应用，这种被称之为“推荐”的在线营销策略，已经证明是信息过载的一种有效解决手段。
总的来说，推荐系统是根据客户的个性化需求、兴趣习惯等信息，将客户感兴趣的商品、怎么写作等相关信息推荐给客户的个性化信息智能处理系统。和搜索引擎相比，它能够通过研究客户的兴趣偏好，进行个性化处理，由系统发现客户的兴趣点，从而引导其发现自己的信息需求。目前这些系统已经广泛应用于电影、音乐、网页等方面的在线营销。
进来，越来越多的推荐系统尝试使用各类先进算法提高其推荐效果；根据基础算法的不同，可将推荐系统分为三类：基于内容的推荐、协同过滤推荐以及组合推荐。目前，协同过滤推荐系统最为流行。协同过滤推荐系统往往采用最近邻匹配算法，利用客户的历史喜好信CACER：一种基于疯狂蚁群算法的电子商务推荐模型由提供海量免费论文范文的www.udooo.com,希望对您的论文写作有帮助.息计算客户之间的距离，然后利用目标客户的最近邻居客户对商品评价的加权评价的绝对值来判断客户对特定商品或怎么写作的喜好程度，从而根据这一喜好程度来对目标客户进行推荐。协同过滤对推荐对象没有特殊的要求，并且能处理非结构化的复杂对象，如音乐、电影，因此在电子商务领域应用较广。
蚁群优化算法（ACO）是一个相对较新的启发式算法，能够有效的进行局部搜索，并解决相应组合优化理由，如推荐等[8]；该算法模拟了自然界中蚁群寻找食物的过程，从而寻找理由的最有效的途径。第一个蚁群算法是由clolrni等人开发，并成功地应用于旅行商理由（TSP）。而bulleneimer等人基于这种基本思路，设计了路径选择优化解决方案，并随之提出了一种改善的蚁群算法。因为推荐理由是非常复杂的，而基本的蚁群算法不能直接应用，并满意的系统的性能需要；因此，许多研究人员都提出了新的策略来提高改善方案来解决一系列不同的理由。
上述研究成果，成功证明了蚁群算法与其他算法来解决各类大规模复杂理由的可能性；基于他们的启发，本文提出了一种新型的电子商务推荐系统模型（CACER： crazy ant colony E-commerce recommendation）。本文其他部分安排如下：第二部分总结传统推荐系统的不足，并提出系统模型；第三部分，给出疯狂蚁群算法及其应用，包括算法数据结构、运转流程等。最后两部分分别给出仿真实验结果，并总结全文。
1 系统模型
基于推荐算法的不同，推荐系统通常可以分为以下三类：①基于内容的推荐系统：此类系统能够根据客户对内容的偏好，自动检索过滤众多内容，选择一些与客户偏好匹配度高的信息推荐给客户，并能够依据相似度的高低顺序，产生推荐商品列表。由于其信息依赖程度较低，因此被应用在NewWeeder、Tapestry、InfoFinder、Mooney等系统中。
②基于协同过滤的推荐系统：此类系统基于一组兴趣相同的客户或项目进行推荐，它根据邻居客户（与目标客户兴趣接近或相似的客户）的偏好信息产生对目标客户的推荐列表，因此适合于客户数量较大的电子商务系统，目前已被应用在Amazon、Linden、Smith、and York、Sarwar： Sarwar、Karypis、Konstan、Riedl、Hofmann等系统中。
③组合推荐系统按不同的需求，采取不同的混合策略将不同推荐类型或推荐算法进行组合并推荐。例如，Adomicius和Tuzhilin等提出了一种基于上下文感知的推荐系统，该系统将传统的“用户-项目”二维评分效用模型拓展为包含多种上下文信息的多维评分效用模型，因此取得了较好的精度和客户满意度。
我们的系统模型基于内容推荐，并根据疯狂蚁群算法进行了改造和创新，系统的整体结构描述如下（图1）。该模型中，应用了部分的传统模块用于实现推荐系统，例如商品的联机事务处理（OLTP）、数据挖掘、商务统计与在线数据库等。整个系统的核心是基于疯狂蚁群算法的内容推荐。
①该系统运营的第一阶段是在客户偏好分析模块中运转的，即：在商品的浏览、订货、购写等过程中的将客户群体细分成不同的集群的推荐处理模块；相关信息保存在电子商务系统的基本信息数据库（交易数据等内容）。
②客户产生的实时信息，包括：客户在线检索和商品搜索选项等，将分类并保存到在线数据库，为后续工作提供交易信息等信息支持。
③B/C OLTP（联机事务处理）和数据挖掘模块集成了电子商务的基本信息和实时交易信息推荐的所有内容，包括客户及相关商品的信息。
④推荐处理模块，该模块是系统的核心，它通过上述模块采集并预处理过后的集成数据，生成推荐列表；疯狂蚁群算法是其中的关键。该算法搜索客户与商品的匹配指数从两个方向上生成推荐列表。首先是根据客户的偏好生成推荐列表；其次是当新商品上架，也可根据客户的需求信息，向客户进行推荐，提供自动选择倡议怎么写作。⑤推荐通信模块，该模块的推荐实施策略包括电子邮件、在线广告、弹出式通知等等，这也是一个双向通信模块，即：它不但可以推荐的商品列表给客户，而且可以跟踪生成新商品的特征词，当客户对新商品进行评价或购写新商品时，该模块可以通过智能推理算法预测的该商品归于于哪一类客户的偏好。
上述模块与本模型，均能够通过较小的修改，嵌入其他推荐系统。而疯狂蚁群推荐算法也可以用于替换其他基于内容的推荐算法，具体细节参见下文。
2 疯狂蚁群推荐算法
本节将算法的主要思想、模型和数据结构以及实现。2.1 主要思想与模型
疯狂蚁群算法是一种特殊的优化算法，它是基于Agent模型（蚂蚁），实现模拟自然界蚂蚁觅食的群体行为，并发展出对应的合作机制，以协助他们共享觅食经验，最终找到食物源之间的最短路径。总的来说，该算法的核心是以客户和商品信息为基础的启发式算法，通过数据挖掘产生的信息反馈，最终实现推荐的优化。其基本步骤如下：①在初始状态下，推荐系统中的蚂蚁根据人工设定进行觅食（寻找匹配度高的客户或商品），此时系统内没有信息素，那么它们各自会随机的选择一条路径。②在下一个状态，每只蚂蚁到达了不同的点（状态或数据点），从初始点到这些点之间留下了信息素，蚂蚁继续爬行，已经到达目标（信息列表满或者客户退出系统）的蚂蚁返回算法模块，并根据信息素，放出下一批蚂蚁，它们都会按照各条路径上信息素的多少选择路线（selection），更倾向于选择信息素多的路径走（当然也有随机性）。③进行再下一个状态，刚刚没有蚂蚁经过的路线上的信息素不同程度的挥发掉了（evaporation）。而刚刚经过了蚂蚁的路线信息素增强（reinforcement）。然后又出动一批蚂蚁，重复第2个步骤。每个状态到下一个状态的变化称为一次迭代，在迭代多次过后，就会有某一条路径上的信息素明显多于其它路径这通常就是一条最优路径。我们的算法可以利用3种蚂蚁Agent处理推荐理由。
第一种是侦察（Scout）蚂蚁，它被派出去收集新推荐商品吸引程度。第二种是承载（Carrier）蚂蚁，它被派出找到客户，并产生推荐商品列表，或者根据新商品，产生推荐客户列表。最后是推销员（Sale）蚂蚁，它主要用于生成随机推荐广告。各种蚂蚁模型描述如下图2所示。
随着客户历史信息的不断丰富，侦察蚂蚁可以从B/C OLTP和数据发掘模块开始着手，进入电子商务基本数据库，通过交易统计模块等提供预处理怎么写作。在他们的搜索过程中，收集和分类新客户的信息。进入电子商务基本数据库后，就可以使用客户购物兴趣等信息素，在各类商品的信息项上记录关键词和描述词（信息素）。最终，这些蚂蚁将回到B/C OLTP和数据挖掘模块，汇总“爬行”过的商品的客户吸引程度；数据发掘模块将由此发现商品和客户之间的关系，及其变CACER：一种基于疯狂蚁群算法的电子商务推荐模型由优秀论文网站www.udooo.com提供,助您写好论文.化趋势；并生成初级的商品推荐目录。
当注册客户进入电子商务系统，承载蚂蚁从推荐处理模块（蚁巢）出发，为注册客户提供全程跟踪怎么写作。它们不但可以收集和发现客户的个性化的需求信息。偏好等信息将被它们收集起来，与商品信息中携带的信息素进行匹配，并选择信息素浓度较高的商品进行更新，并注入电子商务基础数据库。最终，承载蚂蚁的信息浓度达到一定程度后，退回推荐处理模块，根据采集到的信息，向客户推荐相关商品。
当未注册的用户进入系统和浏览上平项目时，推销员（Sale）蚂蚁从推荐处理模块（蚁巢）出发，而未注册客户的的个性化信息均由其收集后进行分类。此外，此类蚂蚁可以收集和发现未注册客户的兴趣点、客户的浏览序列。收集到的偏好等信息，将用于相关信息素的更新；当这些推销员蚂蚁结束处理，进入了电子商务基本数据库，汇总并更新相关商品信息后；这些信息将通过B/C OLTP和数据挖掘模块，优化推荐列表的生成过程和结果。
2.2 匹配算法
因为传统的客户兴趣偏好和商品属性分处不同的分类空间，因此很难彼此匹配。为解决这一理由，CACER系统模型采用了一些新的技术。除采用历史购物记录挖掘外，CACER系统模型构建了“偏好VS属性”二维匹配空间实现实时推荐。这个关键技术，在实施上将推荐过程分为两个阶段，第一几个阶段是客户偏好注入阶段，主要是将客户历史的浏览、购物等活动产生的记录，进行标准化和归类，生成不同的簇注入一个二维空间中，每个结果簇由一定的客户偏好域组成。在第二个阶段是进行购物偏好预测和推荐生成，在此过程中，新商品往往还没有顾客购写，CACER模型将其归入初始簇中，然后随着客户评论以及购写过程中显现出来的偏好或者延误，通过三类蚂蚁的“爬行”进行调整，从而将其分类到特定的簇中。最终，当进行客户推荐时，通过蚂蚁的爬行，该空间中可以生成该客户的偏好预测值，并根据空间中最接近的邻居商品，生成TopN推荐商品列表。图3说明了该流程。在这个实例中，一只蚂蚁带着若干客户兴趣信息进入了该二维空间，它将这些信息素标识在空间中。其他蚂蚁随后陆续到达该空间，对若干商品进行了信息素标识。当两类信息素被标识完毕后，一只销售蚂蚁或者承载蚂蚁从中收集了所需的TopN商品列表，然后将其交给推荐实施模块或直接推荐给未注册客户。
3 实验结果与分析
我们进行了仿真实验，来验证CACE算法和相关系统模型的性能。该实验的配置如下：商品数据集为某电子商务网站提供的500种商品及其描述信息。通过CACER系统前后，客户的点击、订货以及购写情况（随机抽取其中十分之一的商品信息）对系统和算法性能进行了考察。图4A表明，CACER系统模型能过帮助电子商务系统较好的吸引客户的注意力；图中显示，每一个对比项中，使用CACER后，客户都在广告页面上停留了较多时间，说明个性化的广告比较符合客户的兴趣偏好。图4B表明，CACER系统模型使得客户（注册和非注册客户）浏览商品的关注时间增加了12%，体现了该系统模型较好的客户需求预测精度，而且CACE算法的运转效果良好。CACER：一种基于疯狂蚁群算法的电子商务推荐模型相关范文由写论文的好帮手www.udooo.com提供,转载请保留.Ant colony optimization （ACO）is a relatively new meta-heuristic and a very effective local search algorithm for a large number of combinatorial optimization problems such as recommendation[8]. It simulates the behior of ant colonies in nature as they forage for food and find the most efficient routes from their nests to food sources. The first ACO algorithm was developed by Clolrni et al. and succesully applied to the treling salean problem （TSP） based on the path-finding abilities of real ants. Bulleneimer et al. he designed the first ant system for the vehicle routing problem and then proposed an improved ant system algorithm. Because the recommendation problem is very complicated， the basic ACO algorithms cannot directly apply to the problem with satiaction performance need. Many researchers he proposed new methods to improve the original ACO and applied them succesully to a whole range of different problems. It is a trend to combine ACO with other algorithms to solve very large-scale of the recommendation problem. The development of modern algorithms has led to considerable progress， but each ACO algorithm has its own strength and weakness. Therefore， much research has tried to develop the quest for the performance of hybrid algorithms expecting to achieve the effectiveness and efficiency.In this paper， we present a novel E-commerce recommendation model called CACER （crazy ant colony E-commerce recommendation）. The rest of the paper is organized as follows. Partition 2 makes a summary of traditional models and gives a distributed E-commerce recommendation model. In partition 3， a crazy ant colony algorithm is given and its application is proposed. And in the section， the message and data structuresCACER：一种基于疯狂蚁群算法的电子商务推荐模型由优秀论文网站www.udooo.com提供,助您写好论文. of CACER are given. Then its work flow is described in detail. In partition 4， correctness and efficiency of the algorithm are proved. And compared with a traditional， simulation results will be given in partition 5. At last， partition 6 concludes the paper.
II. SYSTEM MODEL
Based on how recommendations are made， recommender systems are usually classified into the following three categories[8-11]：
Content-based recommendations： Recommendations are provided by automatically matching a customer's interests with items' contents. Items that are similar to ones the user preferred in the past are now recommended. Notice that recommendations are made without relying on information provided by other customers， but solely on items’ contents and users’ profiles. NewsWeeder applied this method to build a net-news filtering system. Other applications include Tapestry， InfoFinder， Mooney， and so on.Collaborative filtering： Recommendations are made for items that people with similar tastes and preferences liked in the past. This technique is widely used and is the preferred method for personal recommendation. Many systems， such as Amazon.com： Linden， Smith， and York， Sarwar： Sarwar， Karypis， Konstan， and Riedl， Hofmann， and they he adopted this technique.
Hybrid approaches： These approaches combine collaborative and content-based methods. Fab is a hybrid contentbased， collaborative webpage recommendation system that eliminates many handicaps of the pure versions of either approach. According to the classification scheme proposed by Adomicius & Tuzhilin， hybrid recommendation systems can be divided into four categories： （1） implementing collaborative and content-based methods separately and combining their predictions， （2） incorporating some content-based characteristics into a collaborative approach， （3） incorporating some collaborative characteristics into a content-based approach， and （4） constructing a general， unifying model that incorporates both collaborative and content-based characteristics.
Then our system model is based on Content-based recommendation with crazy ant algorithms. And it can be figured with Fig.1.
In the model， some traditional modules are utilized to implement recommendation systems， such as Product OLTP and Data mine， Commerce Statistic and On-line DB. And the recommendation processing module is the core of the system model with crazy ant colony algorithms.
In the addition， the system work flow can be described as following.
（1） The first phase is a customer interest profiling step， where product bought by a customer in the past is segmented into different clusters by recommendation processing modules. And related information is committed into these sub-DBs of E-commerce basic information DB， including transaction data and so on.
（2） Some real-time information， including customer retrieval and search items， can be classified and sed into the on-line DB， which provides transaction information and so on to the transaction statistics module.
（3） B/C OLTP and Data mine module integrates E-commerce basic information and real-time transaction information into recommendation items， including customer and related product information.
（4） Recommendation processing module， as our system core， integrates all kinds of data. And crazy ant colony algorithms can be executed by it. Furthermore， the algorithm will search customer-product matching metrics. ThCACER：一种基于疯狂蚁群算法的电子商务推荐模型由提供海量免费论文范文的www.udooo.com,希望对您的论文写作有帮助.en the module can provide recommendation lists， including customer items and interested product items. Also according to customer information， it can automatically choose specific methods to recommend. Through E-commerce push communication， related information can be tranerred to the customers.（5） Push communication module， composed of Email， NoCACER：一种基于疯狂蚁群算法的电子商务推荐模型相关论文由www.udooo.com收集,如需论文查抄袭率.te， and so on， can be bidirectional sub-module. Mainly， it can recommend product items to customers. Subsequently， the new product features are presented to it and the preference value that the customer possibly has for the new product can be predicted by the generalization inference ability of intelligent algorithms. And through other modules， the preference prediction values for all new products are ranked and the top-n items with the highest values will be recommended to the customer.
Note that our model can be extended for other recommendation systems with little modification. And also， crazy ant colony algorithm can be utilized to replace other content-based recommendation algorithms in the system models. And the details of them can be described as following sections.
III. CRAZY ANT COLONY ALGORITHM
In the section， the main idea， model， data structures， and implementation models are given for crazy ant colony algorithm.
A. Main Idea and Model
The crazy ant colony algorithm is a particular algorithm of ant colony optimization （ACO） which is based on agents that simulate the natural behior of ants， develop mechanis for cooperation， and assist them in using experience to find the shortest path between a food source and the nest. ACS is a population-based heuristics that enables the exploration of the positive feedback whereas the ants are able to communicate （ants lay pheromone for indirect communication） information concerning food source via an aromatic essence. The ants lay pheromone and heuristic information to mark trails. As the paths are visited by other ants， some of the trails may be reinforced and others paths may be allowed to evaporate. Pheromone trails can be observed via the number of ants passing through the trail. When there are more pheromones on a path， there is larger probability that other ants will use that path， and therefore the pheromone trail on such a path will grow faster and attract more ants to follow （so called positive feedback）. An iterative local search algorithm tries to search the current paths to neighboring paths until a better solution is found. And the main idea of CAC algorithm can be imagined as following：
Our algorithm can utilize three kinds of ant agents to deal with the recommendation problem. The first is ascout ant， which can be dispatched to collect the attracting trends of new recommendation products. The others are a band of carrier ants， which are sent to find recommendation products for customers. And the lasts are some salean ants， which randomly recommend AD products to customer. And all kinds of ant models are described as following and showed in .figure 2.论文资料由论文网www.udooo.com提供,转载请保留地址.The proposed technology flow divides the recommendation process into two phases. The first phase is a customer interest profiling step， where product bought by a customer in the past is standardized and segmented into different clusters in a two-dimensional space. Each resultinCACER：一种基于疯狂蚁群算法的电子商务推荐模型相关范文由写论文的好帮手www.udooo.com提供,转载请保留.g cluster corresponds to an interest area of the customer. In the second phase， the tasks of product preference prediction and recommendation generation are carried out. For a new product that is not purchased by the customer yet， CACER categorizes it into the right cluster first. Then the products included in this cluster， together with their ratings given by the customer explicitly or implicitly， are used the same two-dimensional space. Subsequently， the product attributes are formalized and presented to the two-dimensional space. Then the preference value that the customer possibly has for the product can be predicted by the generalization inference ability of top match algorithm. At last， the preference prediction values for all new products are ranked and the top-n items with the highest values will be recommended to the customer. Figure 3 depicts the processing flows of the technology.
In the example， one ant entered into the two-dimensional space with some customer interests， and it labeled some pheromones in the space. Then other ants arrived in the space with some product attributes as their pheromones. When two kinds of pheromones were labeled in the superposed space， a salean ant or a carrier ant could collect asimilarity top-N product list. Further， it pushed the list to recommendation processing module and recommend the list to customers.
IV. SIMULATION RESULTS
In order to validate the proposed CACE algorithm， a simulation experiment was carried out. The simulation configuration for the experiment was as follows： First， a product dataset was designed by randomly choosing 500 products. The products in this category represent the most popular items that are often bought by customers from Internet. The behior of a typical customer was simulated by randomly assigning 100 products out of the 500 products with ratings between 1 and 10， and the rest with click， se and purchase number. Then doublel experiements he been respectivly executed with CACER and without it.
Figure 4A shows that CACER can make the systme attract more consumers’ attention than before. And in each simulation group， consumers spend more time in AD webs with CACER than without it. The performance of each model is evaluated in terms of recommendation precision， which measures the percentage that products recommended to a customer are actually liked by the customer. And the simulation results show the delays， in which comsumer browse the recommended products.Futhermore， CACER can increase E-commerce turnovers. And the simulation results （in Fig.4 B） show the turnovers he been more about 12% than before. This can be explained by the unique properties exhibited by the CACER model such as user profile with multiple interests， adaptability to customer interest changes and the good robust learning and generalization inference ability exhibited by three kinds of ants.V. CONCLUSION
In this paper， we propose a CACER recommendation system model to help e-commerce websites provide better personalization service for their customers. The experimental results show that the our recommendation model is superior to the before. In addition， the model should be applied to a real e-commerce eCACER：一种基于疯狂蚁群算法的电子商务推荐模型由专注毕业论文与职称论文的www.udooo.com提供,转载请保留.nvironment where customers give their actual ratings to products for validating its practical effect.
REFERENCES：
[1]Deng Zhi-Hong， Wang Zhong-Hui， Zhang Jie. ROBIN： A novel personal recommendation model based on information propagation [J]. EXPERT SYSTEMS WITH APPLICATIONS， 2013，40（13）：5306-5313.
[2]Jing Y， Liu H. Integrating ontology technology with folksonomies for personalized social tag recommendation [J]. APPLIED SOFT COMPUTING，2013，13（8）： 3745-3750.
[3]Zhao Xiangyu， Niu Zhendong， Chen Wei. Interest before liking： Two-step recommendation approaches [J]. KNOWLEDGE-BASED SYSTEMS，2013，48（8）：57-69.
[4]Hernando Antonio， Bobadilla Jesus， Ortega Fernando. Trees for explaining recommendations made through collaborative filtering [J]. INFORMATION SCIENCES，2013，239（8）：1-17.
[5]Lai Chin-Hui， Liu Duen-Ren， Lin Cai-Sin. Novel personal and group-based trust models in collaborative filtering for document recommendation [J]. 2013，239（8）：31-49.
[6]Reimann Jan， Seifert Mirko， Asann Uwe. On the reuse and recommendation of model refactoring specifications [J]. SOFTWARE AND SYSTEMS MODELING，2013，12（3）：579-596.
[7]Lucas Joel P， Luz Nuno Moreno， Maria N. A hybrid recommendation approach for a touri system [J]. EXPERT SYSTEMS WITH APPLICATIONS，2013，40（9）：3532-3550.
[8]Agarwal Deepak， Chen Bee-Chung， Elango Pradheep. Content Recommendation on Web Portals [J]. COMMUNICATIONS OF THE ACM. 2013，56（6）：92-101.
[9]Li Lin， Zhong Luo， Yang Zhenglu. QUBIC： An adaptive approach to query-based recommendation [J]. JOURNAL OF INTELLIGENT INFORMATION SYSTEMS，2013，40（3）：556-587.
[10]Fernandez-Luna Juan M， Huete Juan F， Castells Pablo. Personalization and Recommendation in Information Access [J]. INFORMATION PROCESSING & MANAGEMENT，2013，49（3）：637-639.
[11]Yuan Xiaofang， Lee Ji-Hyun， Kim Sun-Joong. Toward a user-oriented recommendation system for real estate websites[J]. INFORMATION SYSTEMS，2013，38（2）：231-243.