论文部分内容阅读
【目的】针对用户兴趣随时间推移不断变化的问题,利用主题模型及时间节点函数预测用户兴趣。【方法】使用主题模型生成用户兴趣,针对用户的所有兴趣,分别利用多时间节点函数对每个兴趣的每次出现进行加权,用以预测用户兴趣在下一个时间节点的分布情况。【结果】在Sogou搜索日志上,与基于记忆的用户兴趣模型、基于遗忘曲线的用户兴趣度多阶段量化模型进行对比实验,余弦相似度及KL(Kullback-Leibler)距离均表明本文方法能较准确地预测用户兴趣。【局限】仅在Sogou搜索日志上进行实验测试,还需在其他数据集上进一步检验。【结论】充分考虑用户历史数据中每一个时间点可更准确地对用户兴趣进行预测。
【Objective】 In view of the problem that the user’s interest changes constantly with the passage of time, the topic model and the time node function are used to predict the user’s interest. [Method] The topic model is used to generate user interest. For each user’s interest, the multi-time node function is used to weight each occurrence of each interest to predict the distribution of user interest at the next time. 【Result】 The results of Sogou search log were compared with those of memory-based user interest model and forget-curve-based user interest multi-stage quantitative model. The cosine similarity and KL (Kullback-Leibler) distance all showed that this method can be more accurate Predict user interest. [Limitations] Experiments are performed only on the Sogou search log and further verified on other datasets. 【Conclusion】 User interest can be predicted more accurately by fully considering each time point in user history data.