Progresīvās tīmekļa lietojumprogrammas un mākslīgā intelekta izmantošana e-komercijā

Abstract

Darba mērķis ir izstrādāt prototipu “Magento 2” platformai, kas ļauj izmantot push-notifications ar progresīvo tīmekļa lietojumprogrammu (Progressive Web Application) un mākslīgā intelekta produktu ieteikumiem. Autors apspriež alternatīvu veidu, kā saņemt prognozes – bezkoda mākslīgo intelektu -, lai samazinātu izstrādes laiku, izmaksas un sarežģītību. Pētījumā tika izmantoti vairāki instrumenti, tostarp ScandiPWA, Amazon SageMaker un Magento 2. Pētījumā uzsvērta mākslīgā intelekta un progressīvo tīmekļa lietojumprogrammu (Progressive Web Application) nozīme, un ieviestais risinājums ir potenciāli rentabla alternatīva Magento Commerce Adobe Sensei rīkam. Lai gan rezultāti demonstrē, ka Amazon SageMaker prognožu precizitāte nav labākā pārmērīgas pielāgošanas (overfitting) dēļ, risinājums pierāda, ka e-komercijā var izmantot bezkoda mākslīgā intelekta rīkus un pēc modeļa izveides atvieglot modeļa pielāgošanas procesu. Pastāv arī ierobežojumi - nespēja testēt un salīdzināt ScandiPWA rezultātus ar Luma motīvu dažādās servera konfigurācijās. Kā arī atbilstošas apmācības datu kopas izveide bija problemātiska, jo netika atrasti atbilstoši publiski pieejami dati, kas lika izveidot pielāgotu datu kopu no ierobežotiem novērojumiem un ietvēra datu dublēšanu. Turpmākais darbs varētu ietvert SageMaker Canvas testēšanu, izmantojot mazāk sarežģītus algoritmus, piemēram, piegādes datuma paredzēšanu, un īstu lietotāju rīcības datu apkopošanu, lai kvalitatīvāk apmācītu modeli. Diplomdarbs ir uzrakstīts angļu valodā, tā apjoms ir 101 lapaspuse, tajā ir iekļautas 12 figūras, 9 pielikumi ar 7 tabulām un 21 figūru, kā arī 99 literatūras avoti. Atslēgvārdi: Progresīvās tīmekļa lietojumprogrammas (Progressive Web Applications), Mākslīgais intelekts, E-komercija, ScandiPWA, Magento

The problems discussed in the paper involve a lack of flexibility for business owners in managing the cost, relevance and complexity of the AI solutions used for product suggestions. Existing solutions are either not powered by AI creating relevance issue, are too expensive for small or medium businesses creating cost issue or involve sophisticated code for AI model training creating the issue of complexity. In addition, the usage of stores not supporting PWA technology reduces the means of communication with customers. The purpose of the thesis is to create a prototype for Magento 2 allowing to use push notifications attributed to PWA technology together with AI-powered product suggestions. As AI is a complex tool, the thesis discusses alternative ways of receiving predictions – no-code AI - that would decrease development time, cost and complexity of managing the model when the model has already been created. Several tools were used to address the issue, including Magento 2 as an e-commerce platform, ScandiPWA as a storefront for Magento 2, Amazon SageMaker and other smaller Amazon services for AI prediction generation and API support. As a result, ScandiPWA was found to perform better than the Luma theme in three out of four Lighthouse audit variables, but Luma performed better in terms of performance across different devices and website pages. The SageMaker Canvas interface made model management accessible and editable even by non-developers, but setting up all required Amazon Web Services was more complex than anticipated and may require professional assistance. Costs were transparent and could be estimated in advance using Amazon's cost calculator, and cost management was possible by choosing more or less powerful hardware for AI prediction generation. Combining ScandiPWA and Create Magento App proved to be a quick set-up during development. However, ScandiPWA falls short of being a fully PWA theme due to a lack of push notification implementation in the core app. The research emphasizes the importance of AI and PWA, with the implemented solution as a potentially cost-effective alternative to Magento Commerce. ScandiPWA outperforms Luma in three out of four Lighthouse variables, and with push notification support, showcases practical PWA implementation. Even though results demonstrate that the accuracy of Amazon SageMaker predictions is not the best due to overfitting, the solution proves that no-code AI tools can be used in e-commerce and ease the model adjustment process after the model’s creation. Limitations were noticed during the development - inability to test the performance of ScandiPWA compared to the Luma theme on different machines. Amazon SageMaker (no-code AI tool) limitations related to field types and unallowed missing values. Lastly, creating a training dataset was challenging due to a lack of publicly available data, requiring creation of a custom dataset from limited observations and involving duplication of data. Future researches could compare ScandiPWA and Luma themes on different machines or gather real user behaviour data to train the model. Amazon SageMaker Canvas might not be meant for advanced prediction logic. Future work could involve testing SageMaker Canvas on less sophisticated algorithms, such as delivery date or product demand prediction. Bachelor thesis consists of 101 pages including 12 figures and nine appendices with seven tables and 21 figures, and 99 sources of literature. Keywords: Progressive Web Applications, Artificial Intelligence, E-commerce, ScandiPWA, Magento.