Collaborative　filtering　usually　suffers　from　limited　performance　due　to　the　data　sparsity　problem.　Transfer　learning　presents　an　unprecedented　opportunity　to　alleviate　this　issue　through　transfer　useful　knowledge　from　an　auxiliary　domain　to　a　target　domain.　Cluster-level　rating　patterns　transformation　models　have　been　widely　used　due　to　the　loose　restriction　which　does　not　assume　the　source　overlaps　users　and　items　with　the　target.　However,　previous　researches　have　never　investigated　the　relationship　between　the　codebook　scale　in　transfer　learning　and　the　prediction　accuracy　in　the　target　domain.　Moreover,　all　existing　rating　patterns　sharing　models　fix　the　codebook　scale　without　considering　the　data　features　of　the　source　domain.　In　this　paper,　we　propose　a　novel　model,　namely　ACTL,　to　efficiently　and　automatically　discover　the　appropriate　codebook　scale,　which　balances　both　the　computational　cost　and　prediction　accuracy　and　best　matches　the　size　and　features　of　the　source　domain　for　the　cross-domain　recommendation.　The　extensive　experiments　on　real-world　datasets　demonstrate　that　our　algorithms　get　knowledge　gain　from　the　large　source　domain　and　clearly　and　solidly　outperform　the　state-of-the-art　fixed　scale　codebook　transfer　learning　methods.