iDSS Data Mining Presentation Template

Size: px

Start display at page:

Download "iDSS Data Mining Presentation Template"

宛洪伯长孙
4 years ago
Views:

1 Detecting Online Commercial Intention (OCI) Honghua (Kathy) Dai, Zaiqing Nie, Lee Wang, Lingzhi Zhao, Ji-Rong Wen, and Ying Li 6/19/2006 1

2 Agenda Motivations and introduction to OCI (Online Commercial Intention) A machine learning-based approach for OCI detection Experiments Conclusion and future work 6/19/2006 2

3 Motivation Serving ads will be more effective and less annoying, when user has intent to purchase We are interested in detecting web pages / queries that show intention to commit a commercial activity (purchase, rent, bid, or sell ) 6/19/2006 3

4 OCI vs 3 search goal categories Navigational The immediate intent is to reach a particular site Informational The intent is to acquire some information assumed to be present on one or more web pages. Transactional The intent is to perform some web-mediated activity OCI can bee seen as a new dimension of user search goals. Commercial Navigational walmart hotmail Non-Commercial Informational Digital camera San Francisco Transactional / Resource U2 music download Collide lyrics 6/19/2006 4

5 Define the OCI detection problem A binary classification problem OCI: Query/Page -> {Commercial, Non- Commercial} We can derive the commercial sense from a confidence value that ranges from 0 (no commercial intent) to 1 (strong commercial intent) 6/19/2006 5

6 Framework of Detecting Page OCI 6/19/2006 6

7 Keywords selection Select significant and reliable keywords Significance: Frequency: Keyword selection threshold Max{Pr( k C+ ),Pr( k C )} Sig( k) = 2 1 Pr( k C ) + Pr( k C ) Freq ( k ) = Pr( k C C + ) For simplicity we use the same threshold for the two measures in the experiments. + 6/19/2006 7

8 Page feature composition We define two aspects of properties for each keyword in a page p: nit( ki, p) keyword occurrences in inner text nta( k i, p) keyword occurrences in tag attributes As the result, a page p is represented by a feature vector using these two aspects 6/19/2006 8

9 Detecting query OCI Challenges Only few search queries contain explicit commercial indicators, such as buy, price, rent, discount, etc. Search queries are usually short. Solution Enrich query from external resource (search engine) First result page (Query snippets) Top N landing URLs Query classification problem -> page classification problem 6/19/2006 9

10 Search result page and Landing URLs 6/19/

11 Query OCI Detector based on Top N Landing URLs 6/19/

12 Query OCI Detector based on first search result page Build a dedicated search result page classifier for this purpose 6/19/

13 Labeling process We adopted majority vote: 3 human labelers voted for the labels Initial Web pages and queries were randomly selected from our page/query repository. Pages Commercial Non-Commercial Queries Total /19/

14 Experiment Results - Page OCI detector Reach best performance (CF) when keyword selection threshold = 0.1 (using SVM as the classifier) Keyword Selection Threshold Keyword Number CP CR CF CP, CR and CF are the precision, recall and F1 metrics for detecting commercial intent. 6/19/

15 Experiment Results- Query OCI detector Model based on first result page returns best performance CP CR CF Model based on search result page: OCI(FSRPq) Model based on top N landing pages: OCI(TLPq) SVM Model based on top N landing pages: OCI(TLPq) Naïve Average Query OCI Detector Performance 6/19/

16 OCI Distribution among Query Frequency Ranges 80% 70% 60% Percentage 50% 40% 30% 20% 10% 0% All Queries High Mid Low Very Low Single Query Frequency Range Commercial NonCommercial 6/19/

17 Conclusions The notion of OCI (Online Commercial Intention) and the problem of detecting OCI from pages and queries. The framework of building machine learning models to detect OCI based on Web page content. Based on this framework, we build models to detect OCI from search queries. 6/19/

18 Conclusions (cont.) Our framework trains learning models from two types of data sources for a given search query: content of first search result page (query snippets) content of top landing URLs returned by search engine. Experiments showed that the model based on the first search result page achieved better performance. We also discovered an interesting phenomenon that the portion of queries having commercial intention is higher in frequent query sets. 6/19/

19 Future work Utilize search query click through logs Reduce labeling effort Take user online context into consideration in studying user s online intention Detect at which commercial activity phase a user is (research/commit). 6/19/

20 Future Work (Cont.) Detect more detailed commercial intentions in different verticals Traveling intention and preferences. Branding awareness and preferences. Study how specific the user intention is: Halo2 vs video games cheap airline ticket new york to las vegas vs book a flight Study the correlations between conversion rate and user intention. A lot of more interesting research problems! We are HIRING! Contact:KathyDai@microsoft.com 6/19/

21 Thank You for Your Attention! 6/19/

清华大学

清华大学清华大学综合论文训练题目 : 基于网络用户行为分析的传染病发病趋势研究系专姓别 : 计算机科学与技术业 : 计算机科学与技术名 : 许丹青指导教师 : 刘奕群助理研究员 2010 年 6 月 27 日中文摘要近年来, 传染病的传播与流行已