Mashup　has　emerged　as　a　lightweight　way　to　compose　multiple　web　services　and　create　value-added　compositions.　Facing　the　large　amount　of　services,　effective　service　recommendations　are　in　great　need.　Service　recommendations　for　mashup　queries　suffers　from　a　mashup-side　cold-start　problem,　and　traditional　approaches　usually　overcome　this　by　first　applying　topic　models　to　mine　topic　proportions　of　services　and　mashup　queries,　and　then　using　them　for　subsequent　recommendations.　This　solution　overlooks　the　fact　that　usage　record　can　provide　a　feedback　for　text　extraction.　Besides,　traditional　approaches　usually　treat　all　the　usage　records　equally,　and　overlook　the　fact　that　the　service　usage　pattern　is　evolving.　In　this　article,　the　authors　overcome　these　issues　and　propose　an　end-to-end　service　recommendation　algorithm　by　extending　collaborative　topic　regression.　The　result　is　a　generative　process　to　model　the　whole　procedure　of　service　selection;　thus,　usage　can　guide　the　mining　of　text　content,　and　meanwhile,　they　give　time-aware　confidence　levels　to　different　historical　usages.　Experiments　on　the　real-world　Programmable　Web　data　set　show　that　the　proposed　algorithm　gains　an　improvement　of　6.3%　in　terms　of　mAP@50　and　10.6%　in　terms　of　Recall　@50　compared　with　the　state-of-the-art　methods.