This paper proposes a novel admission and replacement technique for web caching, which utilizes the multinomial logistic regression (MLR) as classifier. The MLR model is trained for classifying the web cache's object worthiness. The parameter object worthiness is a polytomous (discrete) variable which depends on the traffic and the object properties. Using worthiness as a key, an adaptive caching model is proposed. Trace driven simulations are used to evaluate the performance of the scheme. Test results show that a properly trained MLR model yields good cache performance in terms of hit ratios and disk space utilization, making the proposed scheme as a viable semi intelligent caching scheme.