基于卷积神经网络的电商数据深度挖掘

发布时间：2018-08-23 08:01

【摘要】：近年来,快速发展的电子商务为人们带来了极大的便利。电商所处的商业环境相较传统商业环境具有更强的动态性与复杂性,这带来了诸多挑战,而数据挖掘技术可以帮助人们更好地应对这些挑战。传统数据挖掘技术无法有效地利用电商中的海量数据,它依赖于耗时、耗力的特征工程,得到的模型可扩展性差。深度学习技术可以有效地利用大量数据,且可以实现自动化地从原始数据中抽取有效特征,具有更高的可用性。在本文中,我们利用深度学习中的卷积神经网络对.电商数据进行挖掘,针对商品搭配推荐与商品销量预测这两个方面,设计了一系列有效的算法及优化方法。具体而言,本文的主要研究内容包括:首先,商品搭配具有广泛应用,如基于用户已购买商品向其推荐可能购买的商品。传统方法通过分析商品历史共同购买记录生成搭配信息,它无法为没有历史购买记录的新商品生成搭配信息。在本文中,我们观察到商家会把商品所有重要属性信息放在标题中,于是设计了一个对拍卷积神经网络对两个商品标题组成的短文本对建模,将文本信息从原始的符号空间映射到特定的样式空间,进而在样式空间中计算两个商品间的搭配程度。其次,商品销量预测对商业决策至关重要,它有助于商家对人力、物力与仓储等诸多方面做出更优的管理。基于时序分析的方法仅能对那些销量变化规律明显的商品做出准确预测;虽然传统机器学习方法可以通过特征工程来考虑更多信息,进而取得更高的准确性,但特征工程限制了模型的可扩展性。在本文中,我们设计了一个新颖的模型,它可以从原始结构化时序数据中通过卷积神经网络自动化提取有效特征,并进一步利用这些特征实现商品销量预测。最后,在真实电商数据集上验证了我们提出算法的有效性。
[Abstract]:In recent years, the rapid development of electronic commerce has brought great convenience to people. The business environment of e-commerce is more dynamic and complex than the traditional business environment, which brings many challenges, and data mining technology can help people to cope with these challenges better. The traditional data mining technology can not make effective use of the massive data in the ecoquotient. It relies on the time-consuming and labor-intensive feature engineering, and the model has poor scalability. Depth learning technology can effectively utilize a large amount of data, and can automatically extract effective features from raw data, so it has higher availability. In this paper, we use convolution neural network pairs in deep learning. Based on the data mining, a series of effective algorithms and optimization methods are designed in this paper. Specifically, the main research contents of this paper are as follows: first, commodity collocation has a wide range of applications, such as based on the user has already purchased goods to recommend the goods they may buy. The traditional method can not generate collocation information for new commodities without historical purchase records by analyzing the historical purchase records of commodities together to generate collocation information. In this paper, we observe that the merchant will put all the important attribute information of the product in the title, so we design a short text pair model of the two commodity titles, which is composed of two commodity titles, and a convolution neural network is designed. The text information is mapped from the original symbol space to the specific style space, and the collocation degree between the two items is calculated in the style space. Secondly, the forecast of commodity sales is very important to business decision, it helps merchants to make better management of manpower, material resources and storage and so on. The method based on time series analysis can only accurately predict those commodities whose sales change is obvious; although the traditional machine learning method can take more information into account through feature engineering, and then achieve higher accuracy. However, feature engineering limits the extensibility of the model. In this paper, we design a novel model, which can automatically extract valid features from the original structured temporal data by convolution neural network, and further utilize these features to predict the sales volume of goods. Finally, the validity of the proposed algorithm is verified on the real data set.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP18;TP311.13

【相似文献】