| With the rapid development of Internet technology,the information resources on the network are complicated and exploding.The Recommendation System(RS)has been continuously researched by scholars in various fields due to its characteristics of information filtering.The traditional recommendation algorithm mines the user's preferences mainly by analyzing the user's historicality,and can achieve excellent results in the scenario where the item candidate pool and the user pool are relatively static.At present,a large number of recommended behaviors are completed through real-time online,which requires the recommendation system to respond to user feedback in a timely manner and make continuous recommendations for a period of time.Traditional recommendation algorithms are difficult to adapt to the dynamics of this online environment,leading to the "Exploration-Exploitation" problem.The Muit-Arm Bandit(MAB)is able to process data dynamically,continuously update the strategy with continuous feedback,and handle the balance of "Exploration-Exploitation" well.Therefore,this paper models the recommendation problem as a MAB problem.The arm in the Muit-Arm Bandit corresponds to the item to be recommended,and the reward corresponds to whether the user clicks on the recommended item.Although MAB has excellent theoretical support and application effects,the existing MAB-based recommendation algorithm still has certain limitations.First,only one item is recommended for each recommendation,which is not in line with the recommended form in daily life.Second,the recommended items only consider their accuracy,ignoring other evaluation indicators.In order to solve the two limitations of the existing MAB-based recommendation algorithm,this paper proposes a recommendation algorithm based on Ranked Bandits,trying to solve these two problems simultaneously in one algorithm.'1.In the traditional MAB algorithm,only one item is recommended for each recommendation,and the Ranked Bandits are introduced into the algorithm design.The Ranked Bandits algorithm maps each ranking position to a separate instantiated MAB algorithm.2.Whether it is the traditional MAB or the Ranked Bandits algorithm,the core idea is to optimize the choice of the ann,that is,only focus on the accuracy of the recommendation,ignoring other indicators.In addition to the recommended accuracy,a good recommendation system needs to pay attention to other evaluation indicators,such as diversity and novelty.Therefore,based on the introduction of the Ranked Bandits algorithm,the design of the algorithm is divided into four sub-modules:standard quantization method,index selection,weighting scheme and MAB algorithm selection.In this paper,the linear weighted summation method is chosen to weight the three indicators of accuracy,diversity and novelty.According to different situations,different weighting schemes are set for these three indicators from the perspective of user and ranking.It is worth noting that,in principle,any MAB algorithm can be applied to the proposed algorithm after instantiation,but different MAB algorithm choices have an impact on the performance of the overall algorithm.Therefore,in order to improve the performance of the overall algorithm,this paper improves the existing CONLINBA algorithm,introduces the deep learning model StackDenoising Auto Encoder(SDAE)instead of LDA model to obtain the item features,and generates a new MAB algorithm for instantiation.Finally,on the two public datasets Last.fm and Delicious,the proposed algorithm based on the Ranked Bandits is experimentally verified.The experimental results show that the instantiation algorithm generated after the introduction of deep learning model SDAE is better than LinUCB,MFLinUCB and CONLINBA algorithm to some extent,which reflects the advantages of SDAE model acquisition characteristics.The overall recommendation algorithm based on Ranked Bandits is to a certain extent.The Ranked Bandits algorithm is better than the accuracy-only algorithm.It shows that the algorithm considering multiple indicators has a positive effect on the recommendation performance,which is in line with the expectation of the algorithm design. |
