{"id":{"repo_id":"uiuc","oai_identifier":"oai:www.ideals.illinois.edu:2142/113184"},"canonical_url":"https://search.dev.ndltd.org/etd/uiuc/oai:www.ideals.illinois.edu:2142/113184","repository":{"repo_id":"uiuc","name":"University of Illinois - Urbana-Champaign","base_url":"https://www.ideals.illinois.edu/oai-pmh"},"display":{"title":"Choice modeling and recommendation optimization in presence of context effects","abstract":"Random Utility choice models are supervised learning tools that can be used to estimate the choice behavior of customers facing multiple options. This is accomplished through assigning a utility value to each option and deriving a choice probability for each option. In the presence of context effects, the utility perceived from individual options is not fixed and depends on other options that are offered beside them. While context effects are well explored in the marketing and psychology literature, very little work has been done on incorporating these effects in revenue management systems and product recommendation modules. In this thesis, we propose three sets of machine learning models in order to capture these effects in different settings with different input data structures. For these settings, we also study combinatorial problems concerned with finding the optimal set of products to offer to the customer including (i) assortment optimization problem or reward maximization problem, (ii) click through rate optimization problem, and (iii) customer surplus optimization problem. The first model we propose is a random utility discrete choice model which captures context effects in sparse choice/click data sets and under single-choice outcome assumption. In the proposed model, the perceived utilities from products are dependent on the whole choice set recommended to the customer, and choice probabilities have Multinomial Logistic Regression-type structure. We show the prediction power of this model by testing it on a relevant real data set and prove the NP-hardness of the assortment optimization problem under the proposed model. Several polynomially solvable special cases of the model are identified that also perform well in our empirical validation for our data set. We obtain some easily verifiable conditions for the monotonicity and submodularity of the assortment optimization objective in order to provide some approximation guarantees. Second, we propose a utility based listwise logistic regression model, which is applicable in estimating the context effects in dense data sets with a multi-choice outcome assumption. We show the predictive and descriptive power of this model through an extensive empirical study on real click data sets chosen from diverse categories of products. We prove the NP-hardness of the Assortment Optimization Problem (AOP) under the general CL model, and show that when some specific types of contextual interactions are dominant in the data, the AOP is tractable. Third, we propose a featurized choice model, in order to capture context effects when the input data is featurized. We study the top-$K$ retrieval problem which focuses on finding $K$ relevant products/documents for a given query. We train a featurized estimator that can measure the context effects among the objects through mapping their features to contextual interaction terms by using an underlying neural net structure. We empirically validate the estimator on a real data set and prove the NP-hardness of the top-$K$ retrieval problem for the proposed model. For all three sets of models, to circumvent NP-hardness, we design heuristic algorithms and test their efficiency through extensive numerical studies. Different models proposed in this thesis and the relevant empirical studies, as well as the recommendation optimization results, shed more light on the contextual behavioral patterns observed in customers' choice behavior in e-commerce platforms, and how to further optimize the recommender systems by considering these patterns. To the best of our knowledge, this thesis is the first systematic study and its findings can help in designing operational recommender systems that capture complex contextual patterns in large scale data sets.","abstract_html":"Random Utility choice models are supervised learning tools that can be used to estimate the choice behavior of customers facing multiple options. This is accomplished through assigning a utility value to each option and deriving a choice probability for each option. In the presence of context effects, the utility perceived from individual options is not fixed and depends on other options that are offered beside them. While context effects are well explored in the marketing and psychology literature, very little work has been done on incorporating these effects in revenue management systems and product recommendation modules. In this thesis, we propose three sets of machine learning models in order to capture these effects in different settings with different input data structures. For these settings, we also study combinatorial problems concerned with finding the optimal set of products to offer to the customer including (i) assortment optimization problem or reward maximization problem, (ii) click through rate optimization problem, and (iii) customer surplus optimization problem. The first model we propose is a random utility discrete choice model which captures context effects in sparse choice/click data sets and under single-choice outcome assumption. In the proposed model, the perceived utilities from products are dependent on the whole choice set recommended to the customer, and choice probabilities have Multinomial Logistic Regression-type structure. We show the prediction power of this model by testing it on a relevant real data set and prove the NP-hardness of the assortment optimization problem under the proposed model. Several polynomially solvable special cases of the model are identified that also perform well in our empirical validation for our data set. We obtain some easily verifiable conditions for the monotonicity and submodularity of the assortment optimization objective in order to provide some approximation guarantees. Second, we propose a utility based listwise logistic regression model, which is applicable in estimating the context effects in dense data sets with a multi-choice outcome assumption. We show the predictive and descriptive power of this model through an extensive empirical study on real click data sets chosen from diverse categories of products. We prove the NP-hardness of the Assortment Optimization Problem (AOP) under the general CL model, and show that when some specific types of contextual interactions are dominant in the data, the AOP is tractable. Third, we propose a featurized choice model, in order to capture context effects when the input data is featurized. We study the top-$K$ retrieval problem which focuses on finding $K$ relevant products/documents for a given query. We train a featurized estimator that can measure the context effects among the objects through mapping their features to contextual interaction terms by using an underlying neural net structure. We empirically validate the estimator on a real data set and prove the NP-hardness of the top-$K$ retrieval problem for the proposed model. For all three sets of models, to circumvent NP-hardness, we design heuristic algorithms and test their efficiency through extensive numerical studies. Different models proposed in this thesis and the relevant empirical studies, as well as the recommendation optimization results, shed more light on the contextual behavioral patterns observed in customers' choice behavior in e-commerce platforms, and how to further optimize the recommender systems by considering these patterns. To the best of our knowledge, this thesis is the first systematic study and its findings can help in designing operational recommender systems that capture complex contextual patterns in large scale data sets.","abstract_has_math":true,"creators":["Yousefi Maragheh, Reza"],"institution":"University of Illinois at Urbana-Champaign","degree_name":"Ph.D.","degree_level":"Dissertation","degree_discipline":"Industrial Engineering","degree_department":null,"school":null,"contributors":["Chen, Xin","Seshdari, Sridhar","Etesami, Rasoul","Zhou, Yuan","Achan, Kannan","Cho, Jason"],"advisors":[],"committee_chairs":[],"committee_members":[],"year":2022,"date_issued":"2022-01-12T22:35:13Z","date_published":"2022-01-12T22:35:13Z","updated_at":"2026-07-22T22:24:53Z","subjects":["context effects","recommender systems","assortment optimization","choice modeling","online retailing"],"languages":["en"],"rights":["Copyright 2021 Reza Yousefi Maragheh"],"rights_urls":[],"identifier_entries":[]},"links":{"outbound_url":"http://hdl.handle.net/2142/113184","outbound_label":"Handle","outbound_source":"dc:identifier"},"metadata_groups":[{"id":"people","label":"People","entries":[{"key":"dc:contributor","label":"Contributor","values":["Chen, Xin","Seshdari, Sridhar","Etesami, Rasoul","Zhou, Yuan","Achan, Kannan","Cho, Jason"]},{"key":"dc:creator","label":"Author","values":["Yousefi Maragheh, Reza"]}]},{"id":"academic_context","label":"Academic Context","entries":[{"key":"dc:date","label":"Dc Date","values":["2022-01-12T22:35:13Z","2024-01-12T22:35:30Z","2021-07-13","2021-08"]},{"key":"dc:type","label":"Dc Type","values":["text","Thesis"]},{"key":"thesis:degree_discipline","label":"Discipline","values":["Industrial Engineering"]},{"key":"thesis:degree_level","label":"Degree Level","values":["Dissertation"]},{"key":"thesis:degree_name","label":"Degree Name","values":["Ph.D."]},{"key":"thesis:institution_name","label":"Thesis Institution Name","values":["University of Illinois at Urbana-Champaign"]}]},{"id":"subjects_keywords","label":"Subjects and Keywords","entries":[{"key":"dc:subject","label":"Dc Subject","values":["context effects","recommender systems","assortment optimization","choice modeling","online retailing"]}]},{"id":"language_rights","label":"Language and Rights","entries":[{"key":"dc:language","label":"Dc Language","values":["en"]},{"key":"dc:rights","label":"Dc Rights","values":["Copyright 2021 Reza Yousefi Maragheh"]}]},{"id":"identifiers","label":"Identifiers","entries":[{"key":"dc:identifier","label":"Identifier","values":["http://hdl.handle.net/2142/113184"]}]},{"id":"additional","label":"Additional Metadata","entries":[{"key":"dc:description","label":"Description","values":["Random Utility choice models are supervised learning tools that can be used to estimate the choice behavior of customers facing multiple options. This is accomplished through assigning a utility value to each option and deriving a choice probability for each option. In the presence of context effects, the utility perceived from individual options is not fixed and depends on other options that are offered beside them. While context effects are well explored in the marketing and psychology literature, very little work has been done on incorporating these effects in revenue management systems and product recommendation modules. In this thesis, we propose three sets of machine learning models in order to capture these effects in different settings with different input data structures. For these settings, we also study combinatorial problems concerned with finding the optimal set of products to offer to the customer including (i) assortment optimization problem or reward maximization problem, (ii) click through rate optimization problem, and (iii) customer surplus optimization problem. The first model we propose is a random utility discrete choice model which captures context effects in sparse choice/click data sets and under single-choice outcome assumption. In the proposed model, the perceived utilities from products are dependent on the whole choice set recommended to the customer, and choice probabilities have Multinomial Logistic Regression-type structure. We show the prediction power of this model by testing it on a relevant real data set and prove the NP-hardness of the assortment optimization problem under the proposed model. Several polynomially solvable special cases of the model are identified that also perform well in our empirical validation for our data set. We obtain some easily verifiable conditions for the monotonicity and submodularity of the assortment optimization objective in order to provide some approximation guarantees. Second, we propose a utility based listwise logistic regression model, which is applicable in estimating the context effects in dense data sets with a multi-choice outcome assumption. We show the predictive and descriptive power of this model through an extensive empirical study on real click data sets chosen from diverse categories of products. We prove the NP-hardness of the Assortment Optimization Problem (AOP) under the general CL model, and show that when some specific types of contextual interactions are dominant in the data, the AOP is tractable. Third, we propose a featurized choice model, in order to capture context effects when the input data is featurized. We study the top-$K$ retrieval problem which focuses on finding $K$ relevant products/documents for a given query. We train a featurized estimator that can measure the context effects among the objects through mapping their features to contextual interaction terms by using an underlying neural net structure. We empirically validate the estimator on a real data set and prove the NP-hardness of the top-$K$ retrieval problem for the proposed model. For all three sets of models, to circumvent NP-hardness, we design heuristic algorithms and test their efficiency through extensive numerical studies. Different models proposed in this thesis and the relevant empirical studies, as well as the recommendation optimization results, shed more light on the contextual behavioral patterns observed in customers' choice behavior in e-commerce platforms, and how to further optimize the recommender systems by considering these patterns. To the best of our knowledge, this thesis is the first systematic study and its findings can help in designing operational recommender systems that capture complex contextual patterns in large scale data sets.","Submission published under a 24 month embargo labeled 'U of I Access', the embargo will last until 2023-08-01","The student, Reza Yousefi Maragheh, accepted the attached license on 2021-07-13 at 12:32.","The student, Reza Yousefi Maragheh, submitted this Dissertation for approval on 2021-07-13 at 12:48.","This Dissertation was approved for publication on 2021-07-13 at 21:10.","DSpace SAF Submission Ingestion Package generated from Vireo submission #16907 on 2022-01-12 at 12:54:50","Made available in DSpace on 2022-01-12T22:35:13Z (GMT). No. of bitstreams: 3 YOUSEFIMARAGHEH-DISSERTATION-2021.pdf: 4931100 bytes, checksum: 282b82c2ce2ec1bff41954f6b18069ae (MD5) LICENSE.txt: 4218 bytes, checksum: 2c4767cfd215edba19698115c5031439 (MD5) PROQUEST_LICENSE.txt: 4564 bytes, checksum: e55b2e26e39bf9eed0017455fc99ad6f (MD5) Previous issue date: 2021-07-13","Embargo set by: Seth Robbins for item 121110 Lift date: 2024-01-12T22:35:30Z Reason: Author requested U of Illinois access only (OA after 2yrs) in Vireo ETD system","Author requested U of Illinois access only (OA after 2yrs) in Vireo ETD system","U of I Only"]},{"key":"dc:format","label":"Dc Format","values":["application/pdf"]},{"key":"dc:title","label":"Title","values":["Choice modeling and recommendation optimization in presence of context effects"]}]}],"canonical_facts":{"dc:contributor":["Chen, Xin","Seshdari, Sridhar","Etesami, Rasoul","Zhou, Yuan","Achan, Kannan","Cho, Jason"],"dc:creator":["Yousefi Maragheh, Reza"],"dc:date":["2022-01-12T22:35:13Z","2024-01-12T22:35:30Z","2021-07-13","2021-08"],"dc:description":["Random Utility choice models are supervised learning tools that can be used to estimate the choice behavior of customers facing multiple options. This is accomplished through assigning a utility value to each option and deriving a choice probability for each option. In the presence of context effects, the utility perceived from individual options is not fixed and depends on other options that are offered beside them. While context effects are well explored in the marketing and psychology literature, very little work has been done on incorporating these effects in revenue management systems and product recommendation modules. In this thesis, we propose three sets of machine learning models in order to capture these effects in different settings with different input data structures. For these settings, we also study combinatorial problems concerned with finding the optimal set of products to offer to the customer including (i) assortment optimization problem or reward maximization problem, (ii) click through rate optimization problem, and (iii) customer surplus optimization problem. The first model we propose is a random utility discrete choice model which captures context effects in sparse choice/click data sets and under single-choice outcome assumption. In the proposed model, the perceived utilities from products are dependent on the whole choice set recommended to the customer, and choice probabilities have Multinomial Logistic Regression-type structure. We show the prediction power of this model by testing it on a relevant real data set and prove the NP-hardness of the assortment optimization problem under the proposed model. Several polynomially solvable special cases of the model are identified that also perform well in our empirical validation for our data set. We obtain some easily verifiable conditions for the monotonicity and submodularity of the assortment optimization objective in order to provide some approximation guarantees. Second, we propose a utility based listwise logistic regression model, which is applicable in estimating the context effects in dense data sets with a multi-choice outcome assumption. We show the predictive and descriptive power of this model through an extensive empirical study on real click data sets chosen from diverse categories of products. We prove the NP-hardness of the Assortment Optimization Problem (AOP) under the general CL model, and show that when some specific types of contextual interactions are dominant in the data, the AOP is tractable. Third, we propose a featurized choice model, in order to capture context effects when the input data is featurized. We study the top-$K$ retrieval problem which focuses on finding $K$ relevant products/documents for a given query. We train a featurized estimator that can measure the context effects among the objects through mapping their features to contextual interaction terms by using an underlying neural net structure. We empirically validate the estimator on a real data set and prove the NP-hardness of the top-$K$ retrieval problem for the proposed model. For all three sets of models, to circumvent NP-hardness, we design heuristic algorithms and test their efficiency through extensive numerical studies. Different models proposed in this thesis and the relevant empirical studies, as well as the recommendation optimization results, shed more light on the contextual behavioral patterns observed in customers' choice behavior in e-commerce platforms, and how to further optimize the recommender systems by considering these patterns. To the best of our knowledge, this thesis is the first systematic study and its findings can help in designing operational recommender systems that capture complex contextual patterns in large scale data sets.","Submission published under a 24 month embargo labeled 'U of I Access', the embargo will last until 2023-08-01","The student, Reza Yousefi Maragheh, accepted the attached license on 2021-07-13 at 12:32.","The student, Reza Yousefi Maragheh, submitted this Dissertation for approval on 2021-07-13 at 12:48.","This Dissertation was approved for publication on 2021-07-13 at 21:10.","DSpace SAF Submission Ingestion Package generated from Vireo submission #16907 on 2022-01-12 at 12:54:50","Made available in DSpace on 2022-01-12T22:35:13Z (GMT). No. of bitstreams: 3 YOUSEFIMARAGHEH-DISSERTATION-2021.pdf: 4931100 bytes, checksum: 282b82c2ce2ec1bff41954f6b18069ae (MD5) LICENSE.txt: 4218 bytes, checksum: 2c4767cfd215edba19698115c5031439 (MD5) PROQUEST_LICENSE.txt: 4564 bytes, checksum: e55b2e26e39bf9eed0017455fc99ad6f (MD5) Previous issue date: 2021-07-13","Embargo set by: Seth Robbins for item 121110 Lift date: 2024-01-12T22:35:30Z Reason: Author requested U of Illinois access only (OA after 2yrs) in Vireo ETD system","Author requested U of Illinois access only (OA after 2yrs) in Vireo ETD system","U of I Only"],"dc:format":["application/pdf"],"dc:identifier":["http://hdl.handle.net/2142/113184"],"dc:language":["en"],"dc:rights":["Copyright 2021 Reza Yousefi Maragheh"],"dc:subject":["context effects","recommender systems","assortment optimization","choice modeling","online retailing"],"dc:title":["Choice modeling and recommendation optimization in presence of context effects"],"dc:type":["text","Thesis"],"thesis:degree_discipline":["Industrial Engineering"],"thesis:degree_level":["Dissertation"],"thesis:degree_name":["Ph.D."],"thesis:institution_name":["University of Illinois at Urbana-Champaign"]},"updated_at":"2026-07-22T22:24:53Z"}