Deep Data Mining In E-commerce Based On Convolutional Neural Networks

In recent years,the E-commerce develops rapidly and has brought great convenience for people.While dynamic and complex business environment in E-commerce raises great challenges,data mining techniques help to overcome these challenges.However,traditional data mining techniques cannot effectively utilize the massive amount of avail-able data in E-commerce and the models constructed by them lack of practicability.The reason is that they rely on the manual feature engineering,which is usually a difficult,time-consuming task and requires expert knowledge.Deep learning models can make full use of these data by extracting effective features automatically from the raw data and thus have a higher practicability.In this paper,focusing on the style match and sales forecast,we design a series of efficient data mining approaches in E-commerce scenarios based on Convolutional Neural Networks.Specifically,the principal studies of this paper are listed as follows:Firstly,style match can be exploited in many commercial applications,such as recommending items to users based on what they have already bought.The traditional frequent item-set mining methods generate match items by analyzing the historical purch-asing patterns and cannot handle new products without historical records.Based on the observation that online sellers will place most of the important attributes of a product in its title description,we design a Siamese Convolutional Neural Network in this paper and feed it with title pairs of items.Those short text pairs will be mapped from the original space of symbolic words into some embedded style space,where the compatibility betwe-en two items is calculated.Secondly,sales forecast has a crucial impact on making informed business decision and can help us to manage the workforce,cash flow and resources etc.Traditional sales forecast is based on time series analysis of historical sales and can only handle well the commodities with stable or seasonal sales trend.Though more recent learning-based methods improve the forecast accuracy by capturing more information in the models,they require case-by-case manual feature engineering and thus are limited in their applicability.In this paper,we design a novel approach to learn effective features automatically from the structured time series data using the Convolutional Neural Network.When fed with raw log data,our approach can automatic-ally extract effective features from that and then forecast sales using those extracted features.Finally,we test our approaches on several large real-world datasets in E-comm-erce and the experimental results validate the effectiveness of our approaches.