Research On Cold Start Problem In Computational Advertisement

The principle of computational advertising is to look for a match between userand advertisement. We can consider the problem as a recommendation problem, werecommend advertisements to users. The trouble is that low CTR(click through rate)in computational advertising lead to poor effect when using traditionalrecommendation algorithm. This problem is also known as cold start problem inrecommendation. In computational advertising, there are two directions to tacklethis problem, the one is formalizing the problem as a multi-armed bandit problem,which can be solved through methods based on reinforcement learning?the otheris content based filter algorithm in recommender system.This paper study the problem of cold start problem in computationaladvertising. We propose two methods, in the first method, we formalize the problemas a session based multi-armed bandit problem by leveraging history search recordof user in the scene of AdWords, we add uncertainty factor to the result of the modelthis method can well balance exploration and Exploitation. In the second method,we propose hash based methods to recommend in cold start phase. We have twoalgorithms, the one is maximal entropy based hash method, we both keep the hashvalue preserving the preference of user over item and the entropy of hash valuemaximization, which can balance precision and recall; the other is boosting basedhash method, we get the hash values bit by bit through boosting method, which leadto strong ability of generalization in the case of long hash value.We also validate the effective of our method. Experiment shows that sessionbandit method outperform the method of directly using the result of model and themethod of no use of users' history search records. The proposed hash basedrecommendation method far exceed other cold start recommendation methods onefficiency, but also outperform some other method on effect.