Traditional collaborative filtering algorithm can not perform well under the condition of cold start. To solve this problem, IPSS-based (Inverse Item Frequence-based Proximity-Significance-Singularity) Item Collaborative Filtering (ICF_IPSS) was proposed, whose core was a novel similarity measure. The measure was composed of the rating similarity and the structure similarity. The difference between the ratings of two items, the difference between the item rating and the median value, and the difference between the rating value and the average rating value of other items were taken into account in the rating similarity. The structure similarity defined the ⅡF (Inverse Item Frequence) coefficient which fully reflected common-rating ratio and punished active users. Experiments were executed on Movie Lens and Jester data sets to testify the accuracy of the ICF_IPSS. In Movie Lens data set, when the nearest neighbor number was 10, the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) was 3.06%, 1.20% lower than ICF_JMSD (Jaccard-based Mean Square Difference-based Item Collaborative Filtering) respectively. When the recommendation item number was 10, the precision and recall was 67.79%, 67.86% higher than ICF_JMSD respectively. The experimental results show that ICF_IPSS is superior to other traditional collaborative filtering algorithms, such as ICF_JMSD.