信息工程与人工智能学院知识库

  随着互联网和数字音乐的迅速发展，各类音乐平台为用户提供了大量的音乐作品。然而随着音乐作品数量的急剧增加，用户面对大量的歌曲信息，很难快速找到自己感兴趣的音乐。为了给用户提供良好的使用体验，同时增加用户对音乐平台的满意度，各类音乐平台使用推荐系统来为用户提供个性化推荐服务。由于用户的兴趣是不断变化的，且随着时间的推移，会出现遗忘现象，进而对用户当前兴趣产生影响。然而常见的推荐系统进行个性化推荐时，很少考虑时间因素，为了使用户在数据庞大的音乐数据里找到自己感兴趣的音乐作品，本文将遗忘现象对兴趣产生的影响纳入个性化的音乐推荐算法中，以提高推荐质量。

  本文考虑到时间因素的影响，提出了两种基于时间衰减函数的音乐推荐模型，一是基于时间衰减函数的协同过滤音乐推荐模型：首先以艾宾浩斯遗忘曲线为基础，拟合指数型时间衰减函数与幂函数型时间衰减函数，其次根据用户听取歌曲频数的分布，建立了合理的评分机制，然后根据修正余弦相似度公式计算评分相似度、歌曲相似度，将两者融合得到歌曲的综合相似度，再引入时间衰减函数得到音乐综合相似度，进行评分预测；二是基于时间衰减函数的LightGBM音乐推荐模型：首先通过时间衰减函数对用户评分进行衰减修正，然后运用LightGBM模型进行评分预测，当评分大于等于阈值时进行推荐。

  在公开音乐数据集Last.fm上，通过实验对提出的两种基于时间衰减函数的个性化音乐推荐算法进行评估。实验表明，引入了时间衰减函数的协同过滤推荐算法（TDF-CF）优于传统的协同过滤算法（CF），引入幂函数型时间衰减函数的协同过滤推荐算法效果更好；引入了时间衰减函数的LightGBM音乐推荐算法（TDF-LGBM）的推荐效果优于未引入时间效应的传统的协同过滤算法（CF），并且引入幂函数型时间衰减函数的LightGBM音乐推荐算法效果更好。最后两种模型的实验结果进行对比分析表明，融入时间效应的TDF-LGBM音乐推荐算法的推荐效果优于TDF-CF的音乐推荐算法的推荐效果，且幂函数型的遗忘曲线对于推荐更有优势，引入幂函数型衰减函数的TDF-LGBM算法的推荐结果最佳。因此，结合时间效应与LightGBM算法建模音乐推荐，能够提高音乐推荐的准确性，为目标用户提供更符合其偏好的音乐作品。

With the fast development of the Internet and digital music, various music platforms provide users with a large number of music works. However, with the quick increase in the number of music works, users face a large amount of song information, and it is difficult for users to fast find the music they are interested in. In order to provide users with a good experience and increase user satisfaction with music platforms, various music platforms use recommendation systems to provide users with personalized recommendation services. Because the user's interest is constantly changing, and over time, there will be a forgetting phe- nomenon, which in turn affects the user's current interest. However, when common recommendation systems make personalized recom- mendations, they rarely consider the time factor. In order to enable users to find music works they are interested in in the huge music data, this article incorporates the impact of forgetting on interests into personalized music recommendations. Algorithm to improve the quality of recommendations.

The paper takes into consideration the impact of time components, and proposes two music recommendation models based on time decay functions. One is a collaborative filtering music recommendation model based on time decay functions: first, based on the Ebbinghaus forgetting curve, fitting an exponential type time decay function and power function time decay function. Secondly, a reasonable scoring mechanism is established according to the distribution of the frequency of the songs listened to by users, and then the scoring similarity and song similarity are calculated according to the modified cosine similarity formula, and the two are merged to obtain the song’s Comprehensive similarity, and then introduce the time decay function to obtain the comprehensive similarity of music, and then make the score prediction; the second is the LightGBM music recommendation model based on the time decay function: firstly, the user's score is attenuated and corrected by the time decay function, and then the LightGBM model is used to predict the score , Recommend when the score is greater than or equal to the threshold.On the public music data set Last.fm, the two proposed personalized music recommendation algorithms based on time decay function are evaluated through experiments. Experiments show that in model 1, the collaborative filtering recommendation algorithm (TDF-CF) with the introduction of a time decay function is better than the traditional collaborative filtering algorithm (CF), and the collaborative filtering recommendation algorithm with the introduction of a power function-type time decay function is better . In the second model, the recommendation effect of LightGBM music recommendation algorithm (TDF-LGBM) with time decay function is better than that of music recommendation algorithm (CF) without time effect, and LightGBM music recom- mendation algorithm with power function time decay function is introduced Better results. Finally, the experimental results of the two models are compared and analyzed. The recommendation effect of the TDF-LGBM music recommendation algorithm incorporating the time effect is better than that of the TDF-CF music recommendation algorithm, and the power function type forgetting curve is more advantageous for recommendation.

On the public music data set Last.fm, the two proposed personalized music recommendation algorithms based on time decay function are evaluated through experiments. Experiments show that the collaborative filtering recommendation algorithm (TDF-CF) that introduces the time decay function is better than the traditional collaborative filtering algorithm (CF), and the collaborative filtering recommendation algorithm that introduces the power function time decay function is better; the time decay function is introduced The recommendation effect of the LightGBM music recommendation algorithm (TDF-LGBM) is better than that of the traditional collaborative filtering algorithm (CF) that does not introduce time effects, and the LightGBM music recommendation algorithm that introduces a power function time decay function is better. The comparative analysis of the experimental results of the last two models shows that the recommendation effect of the TDF-LGBM music recommendation algorithm incorporating the time effect is better than that of the TDF-CF music recommendation algorithm, and the power function type forgetting curve is more advantageous for recommendation. The recommendation result of the TDF-LGBM algorithm that introduces the power function decay function is the best. Therefore, combining time effect and LightGBM algorithm modeling music recommendation can improve the accuracy of music recommendation and provide target users with music works that are more in line with their preferences.