隨著近年來智慧型手機的普及與行動廣告產業的快速發展，三方互利的行動廣告概念架構，使電信業者的寬頻用戶數增加，行動廣告商的顧客數目增加，且使用者能取得網路使用費的折扣或商品折價券，並且利用地點因素和傳統檢索模型做廣告配置[4]。然而，在[4]所提出的廣告配置演算法中所使用的vector space model，卻有嚴重的vocabulary mismatch問題，導致檢索效果低落。 行動廣告推薦具有文件/查詢詞很短，以及須考慮地理位置的因素，如何找到適合的檢索模型與以上兩因素結合將是關鍵的議題。因此，我們區分出對地點敏感的查詢詞及與地點無關的查詢詞，以便使廣告配置演算法能針對各式各樣的查詢詞作不同的處理。針對對地點敏感的查詢詞，如何在廣告配置演算法中，於廣告相關性和廣告距離之間做正確的取捨，將是非常關鍵的議題；針對與地點無關的查詢詞，由於短的文件和查詢詞長久以來已經是行動廣告與生俱來存在的因素，故行動廣告推薦系統仍需解決vocabulary mismatch問題。因此，在此篇研究，我們主要聚焦在利用relevance model及LDA smoothing來舒緩vocabulary mismatch的問題，以改善檢索效果。 在實驗部分，我們比較relevance model, vector space model, query likelihood model with LDA smoothing三者之間的檢索效果，針對34名受試者的平均結果顯示，我們提出利用relevance model的方法達到最好的檢索效果。 雖然relevance model有很好的檢索效果，但其計算時間卻也相當驚人。為了縮短計算時間，在預估relevance model時，我們只把PMLE(w|D)及PQL(q|D)權重值高的document包含進去，而在做KL-divergence的加總時，我們只把P(w|Q)權重值高的words包含進去，這些效率上的考量讓使用relevance model的廣告配置演算法能在短時間內完成複雜的計算。 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.
