融合热度和隐特征分析模型推荐个性化森林旅游景点

Personalized Recommendation of Forest Tourist Attractions Based on Heat Model and Latent Feature Analysis Model

基于采集自雅虎图片分享网站Flickr上带有地理标签的森林旅游照片数据,构建游客-景点关系矩阵,运用隐特征分析模型和旅游景点热度分析模型相融合的方法,分析游客对未去森林旅游景点的感兴趣程度,为游客提供一份专属的个性化的森林旅游景点推荐方案。研究结果表明:对于隐特征分析模型的森林景点推荐,正则化参数有效降低森林景点数据稀疏所导致的预测精度下降,同时合理的特征维度有助于提升森林景点评分预测的精度。此外,融合森林景点热度分析的个性化推荐对推荐准确度和新颖性的提升都有积极帮助。因此,提出加强森林旅游个性化推荐算法的优化以及增强森林旅游个性化推荐的新颖性研究的建议。

⑴Background——With the vigorous development of Internet economy represented by mobile interconnection and the increasing demand for tourism service quality,how to customize individual tourist attractions for tourists has become a hot research topic in the field of tourism service.Recommendation system,as one of important data mining tools,has attracted great attention in travel recommendation,which can predict the next behavior of target users according to their historical behaviors.⑵Methods——Based on the data of forest tourism photos with geographical labels collected from website Flickr,this paper proposes a personalized recommendation method for tourist attractions.Firstly,the relationship matrix between tourists and tourist attractions is established by using the collected data.Then,the latent feature model(LFM)is used to reduce the dimension of data and predict the score of unvisited tourist attractions.After that,a heat model(HM)is proposed to analysis the hot degree of tourist attractions.Finally,a personalized recommendation list is generated for each tourist by integrating the LFM model and HM model.⑶Results——Experiments on 147,612 users of 48,794 tourist attractions show that the prediction accuracy of tourist attractions recommendation algorithm based on the LFM is sensitive to the parameters setting,especially the regularization term.The average visited forest attractions of every tourist is 46.99,while the average visitors for each attraction is 141.95.For different values of regularization term(i.e.,2-1 to 2-10),the RMSE value of the LFM varies from 0.92 to 1.09.The same phenomenon occurs in the convergence speed of LFM in terms of iterations.For example,the best RMSE can be achieved after 50 iterations whenλ=2-1,while the number of iterations is 100 withλ=2-3.Furthermore,the size of latent feature dimension also effects the performance of the LFM.When the size of latent feature dimension changes from 20 to 100,the RMSE fluctuation of LFM is obvious(i.e.,from 0.923 to 0.924).Finally,we evaluate the effect of integrating HM and LFM into a tourist attraction recommendation model.The results show that the performance of LFM model with HM is better than using LFM alone in terms of recommendation precision and novelty.Specifically,compared with using LFM alone,when the length of recommendation list is 20,the accuracy of recommendation increases by 36%.⑷Conclusions and Discussions——The results demonstrate that the forest tourist attractions data with sparsity and high-dimensional features can be well solved by adjusting the regularization parameters of LFM and designing reasonable feature dimension.Furthermore,most existing methods for tourist attractions recommendation focus on the prediction accuracy.LFM with HM model can effectively improve the novelty of recommendation without loss of recommendation precision.Therefore,we suggest that the novelty of personalized recommendation in forest tourism should be strengthened.For example,the life cycle of forest tourist attractions should be considered when designing the forest tourist recommendation methods.Meanwhile,the optimization and improvement method for the forest tourism recommendation solutions will be an important research topic for forest tourism in the future.For instances,in view of the sparse and high-dimensional features of forest tourism data,the sensitivity of model parameters to recommendation results should be fully considered,and the parameters of the model should be optimized by integrating network search or heuristic intelligent algorithms.