Research On Collaborative Filtering Recommender Algorithm Based On Time Effect

Traditional collaborative filtering algorithm does not take into account the time effect of user interest, ignoring the time information that user behavior occurs. In real life, user interest may change over time, and user may show different interest patterns in different time contexts. Studying the time effect of user interest is of great significance to improve the ability of recommender algorithm to predict user's current interest. In recent years, some studies have proved that using time effect can effectively improve the accuracy of collaborative filtering recommender algorithm. At present, the time effect based collaborative filtering recommender algorithms are mainly divided into two categories:time-aware algorithm and time-dependent algorithm. Time-aware algorithm regards time as a context, and mainly researches the cycle repetition of user's interest, arguing that user will show same interest pattern in same time context. The time-dependent algorithm considers time as a continuous variable, and focuses on the time-decay nature of user interest, arguing that user interest will change over time. The work of this dissertation is focused on the time effect of collaborative filtering recommender algorithm and the main contributions of this dissertation are as follows:(1) Through an in-depth study on the use of contextual pre-filtering technology to implement time-aware algorithm, the limitation only using a single time classification of user micro-profiling technology is found. A combination of different time classifications of contextual pre-filtering algorithm is proposed in this dissertation. This new algorithm can take advantage of multiple user interest patterns in different time contexts, as well as more user behavior data.(2) Through an in-depth study on the use of contextual pre-filtering technology to implement time-aware algorithm, the deficiency of the rating mapping algorithm used in user interest modeling phase is found, an improved rating mapping algorithm is proposed in this dissertation.(3) Through an in-depth study on the implementation techniques of time-dependent algorithm, a time period pre-filtering technique based on implicit feedback is proposed. This technique divides the user behavior data into several different time period sub data sets according to the chronological order, and uses these sub data sets to predict user's current interest. By assigning different weights to the results obtained through using different sub data sets, the importance of user behavior data in different time periods for predicting user's current interest is distinguished. Finally, the final rating prediction is calculated by the weighted sum of predicted ratings using different time period sub data sets.(4) Based on the previous work, a new algorithm combining time-awareness with time-dependence is proposed in this dissertation. This algorithm uses pre-filtering technique, considering both time cycle repetition and time decay effects of user interest. This algorithm divides user behavior data into different sub data sets according to different time classifications and different time periods. The multiplication of corresponding time correlation weight and time period weight is used as the new weight for the predicted rating obtained by using different sub data sets, and finally the final rating prediction is to calculate the weighted sum of predicted ratings obtained by using the different sub data sets using the new weight.By conducting some comparative experiments on the public Last.fm dataset, it is proved that the algorithms proposed in this dissertation can further improve the accuracy of collaborative filtering recommender algorithm.