改善三贏架構之行動廣告效果

摘要

Mobile ad recommendation features short documents/queries and location factor. How to combine these two factors with proper IR model is the key issue. Hence, we distinguish location-sensitive queries from location-independent queries such that the ad-matching algorithm could react differently for various queries. For those location-sensitive queries, it is difficult to attain excellent recommendation effectiveness with traditional retrieval model. The relevance of recommended ads under the constraint of user accessibility could not be as relevant as the relevance of recommended ads without limit of distance. Thus, to model the tradeoff between ad relevance and ad distance to the user in personalized ad-matching algorithm is a key issue. As for location-independent queries, mobile ad matching system still have to solve vocabulary mismatch problem because short documents and short queries have been the inherent part of mobile advertising. Hence, we focus on improvement of retrieval effectiveness to relieve vocabulary mismatch problem via relevance model and LDA smoothing.We compare the retrieval effectiveness of the relevance model with vector space model and query likelihood model with LDA smoothing. The average result over 34 users showed that our proposed approach with relevance model achieved the best performance.Although relevance model is effective in ad matching, the computation time takes long time due to complicated calculation. Thus, we consider only high-weighted PMLE(w|D) values and high-weighted PQL(q|D) values in the estimation of relevance model, and top-ranked words with high-weighted P(w|Q) value in the summation of KL-divergence. The efficiency consideration makes the complex computation of the ad matching algorithm with relevance model finished in limited time.