Academia.eduAcademia.edu

Figure 3 – uploaded by Ajit Balakrishnan

See full PDFdownloadDownload figure
Figure 3: Comparison of performance of proposed recommendation system and category-recency based recommendation System According to Herlocker et al. [9], evaluating recommender systems and their algorithms is inherently difficult because different algorithms perform differently on different data sets. Also, such algorithms may be entirely inappropriate in a domain where there are many more items than users or vice versa. Similar differences exist for ratings density, ratings scale, and _ other properties of data sets. Our goal was to compare performance of the proposed recommendation system over previously existing system which suggests items from same category, recency- wise. Additionally, we wanted to expose a good variety of items to users rather then showing few items in random fashion. To meet our purpose, we used method of split testing (also called A/B testing), where half the time recommendations are served through previous system and rest of the time, recommendations are presented via proposed system.

Figure 3 Comparison of performance of proposed recommendation system and category-recency based recommendation System According to Herlocker et al. [9], evaluating recommender systems and their algorithms is inherently difficult because different algorithms perform differently on different data sets. Also, such algorithms may be entirely inappropriate in a domain where there are many more items than users or vice versa. Similar differences exist for ratings density, ratings scale, and _ other properties of data sets. Our goal was to compare performance of the proposed recommendation system over previously existing system which suggests items from same category, recency- wise. Additionally, we wanted to expose a good variety of items to users rather then showing few items in random fashion. To meet our purpose, we used method of split testing (also called A/B testing), where half the time recommendations are served through previous system and rest of the time, recommendations are presented via proposed system.

Related Figures (3)

through automatic tool, tag selector module selects few important tags among these raw tags. Selected tags are then queried on search engine to get relevant set of items. On the other side, we have information regarding items and logs for users’ activities. The component called data analyzer does statistical analysis on the logs and find co-occurrence patterns to suggest another set of items. Data analyzer also analyses the information regarding community opinions for the items that are outcome of the search and co- occurrence analysis. Finally, recommendation system finds ranked recommendations based on search relevance score, community score and co- occurrence score. Figure 1: Architectural components of generic recommendation system
through automatic tool, tag selector module selects few important tags among these raw tags. Selected tags are then queried on search engine to get relevant set of items. On the other side, we have information regarding items and logs for users’ activities. The component called data analyzer does statistical analysis on the logs and find co-occurrence patterns to suggest another set of items. Data analyzer also analyses the information regarding community opinions for the items that are outcome of the search and co- occurrence analysis. Finally, recommendation system finds ranked recommendations based on search relevance score, community score and co- occurrence score. Figure 1: Architectural components of generic recommendation system
Figure 2: Dynamic scoring function that handles community bias Community opinion can be captured in various ways depending on the application. In case of video, music and photo, we do it through rated up, rated down, number of views, frequency of sharing, etc. In case of news articles, we find number of time an article is viewed, forwarded to others, commented by users, etc. Similarly, for books and e-shopping products, we look at number of instances sold, view counts, how many times added to shopping cart, etc. Our past experience and literature study reveal that giving more emphasis to such community feedbacks every time, tend to accumulate to the score. As new visitors also view and rate the same items, the community score is boosted again and again.
Figure 2: Dynamic scoring function that handles community bias Community opinion can be captured in various ways depending on the application. In case of video, music and photo, we do it through rated up, rated down, number of views, frequency of sharing, etc. In case of news articles, we find number of time an article is viewed, forwarded to others, commented by users, etc. Similarly, for books and e-shopping products, we look at number of instances sold, view counts, how many times added to shopping cart, etc. Our past experience and literature study reveal that giving more emphasis to such community feedbacks every time, tend to accumulate to the score. As new visitors also view and rate the same items, the community score is boosted again and again.
Figure 4: Comparison of unique items consumed in proposed system and category- recency based system number of days and found that results of split testing are in favor of proposed recommendation system. The results for videos, photos and music are shown in Figure 3. We also found that number of unique media (video, photo, music) consumed by users over testing period, has also gone up significantly especially in case of photos and videos as shown in Figure 4. Thus, proposed
Figure 4: Comparison of unique items consumed in proposed system and category- recency based system number of days and found that results of split testing are in favor of proposed recommendation system. The results for videos, photos and music are shown in Figure 3. We also found that number of unique media (video, photo, music) consumed by users over testing period, has also gone up significantly especially in case of photos and videos as shown in Figure 4. Thus, proposed
Related topics:
Computer ScienceCollective IntelligenceRecommender System
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved