Due　to　the　sparse　initial　interaction　data,　the　cold-start　problem　poses　a　challenge　to　recommendation　networks,　e-commerce　websites,　and　video　platforms,　etc.　Traditional　user　profiling　can　predict　preferences　but　raises　privacy　concerns.　To　address　this　issue,　this　paper　proposes　a　cross-domain　knowledge-shared　transfer　learning　model　that　is　initially　trained　on　a　cross-domain　mixed　dataset　to　allow　multi-task　learning.　This　shared　knowledge　enhances　the　generalizability　of　the　model　and　minimizes　the　dependence　on　sparse　data.　Different　from　traditional　recommender　systems,　this　model　can　utilize　limited　user-item　interaction　data　to　make　accurate　recommendations　and　reduce　the　cold-start　problem.　After　initial　training,　the　model　is　fine-tuned　on　the　target　domain　data.　To　avoid　overfitting　and　inefficiency,　an　extended　Transformer　module　is　used,　extracting　multi-scale　hidden　states　and　a　Transformer　encoder,　which　are　updated　at　each　time　step.　Experimental　results　on　the　Amazon　dataset　show　that　the　model　can　adapt　to　new　tasks　more　efficiently,　reducing　computational　cost　and　risk　of　overfitting.　©　2023　IEEE.