A core of web caching is the replacement policy of web cache. first, we collect the keywords which represent user's interest from menus of the IPTV system. then, joined these keywoords, content types and the access characteristics, this paper proposes a new replacement algorithm named GDSF-AI (greedy-dual-size-frequency-access-interest) to ensure the documents in the cache system are most interested ones for the users. compared with traditional algorithm, this policy is at least 3.0% better than others in hit rate is at least 1.5% better in byte hit rate.