基于用户浏览日志的移动购买预测研究

Predicting Mobile Purchase Decisions Based on User Browsing Logs

【目的】对用户在移动购物APP进行的信息浏览与商品购买行为特征进行描述分析,并尝试预测商品购买。【方法】在日志请求参数与用户信息行为类型之间建立映射,得到用户的行为解析,进一步分析用户行为特征后,使用Logistic二元回归和C&R决策树两种方法建立商品支付购买预测模型。实验数据来自2015年3月某移动购物APP的290位重度用户的3 923 429条服务器端日志。【结果】在用户浏览行为特征方面,用户周内使用平稳,夜晚睡前达到高峰;最关注单品详情,浏览程度越深越有可能查看店铺信息并进行商品、店铺信息的分享;用户对商品的浏览呈幂律分布,90%的商品浏览记录都在16条以下。在用户购买行为特征方面,有9条浏览记录的商品、提交了订单的商品最有可能被购买;商品购买与浏览、分享单品和店铺信息次数呈正相关。在商品支付购买预测方面,C&R决策树预测准确率稍高于Logistic二元回归,然而变量种类远少于后者。【局限】日志数据可能不能准确反映用户的操作行为;对于用户行为的解析有一定模糊性;数据来自重度用户,可能不具有普适性;数据来自于3月份这个时间段,可能会受前后浏览或购买行为的影响。【结论】用户浏览及购买行为特征可帮助移动购物APP完善产品功能,提升用户体验;Logistic二元回归相比于C&R决策树可以更好地预测商品支付购买。

[Objective] This research characterizes users＇ browsing patterns, aiming to predict their purchasing decisions on mobile shopping applications. [Methods] First, we mapped the request parameters of the logs with users＇ information behavior types. Then, we used logistic binary regression and CR decision tree techniques to establish models to predict the buying decisions. The data set included 3,923,429 lines of server logs generated by 290 heavy users of a popular mobile shopping app in March 2015. [Results] We found that the frequency of users＇ browsing behaviors was stable during the weekdays and reached its peak every night before bedtime. Users paid much attention to product details and those with deeper browsing behaviors are more likely to read introduction to the shop and share related information. The number of views was in line with the power-law distribution and 90% of the merchandise was checked less than 16 times. We also found that goods viewed by 9 times and placed in the carts were most likely to be bought. There was a positive correlation between the purchases of goods and the numbers of views or sharing of the item and the shop. The CR decision tree model＇s prediction accuracy was slightly higher than that of the Logistic binary regression model. However, the former＇s variable types were far less than the latter. [Limitations] Logs cannot fully reflect all users＇ behaviors, which lead to some ambiguity of our analysis. The conclusion might not tell the whole story since the logs were generated by heavy users in one month. [Conclusions] The pattern of user browsing and buying behaviors could be used to enhance their experience of the mobile shopping applications. Logistic binary regression might better predict users＇ buying decisions than the CR decision trees model.