While actions that products enable their buyers to do are considered to play a signi cant role in purchase decisions, little attention is given to the integration of such actions in recommender systems. This work proposes the ActionRec framework, to semantically annotate the actions expressed by buyers in product reviews, and to capture the actions' linked entities in an openly accessible knowledge graph, which is used to develop action-aware recommendation apps. A web-based prototype is developed to demonstrate the potentials of integrating the annotated actions in a simulated product recommendations environment.
Recommender systems have been gaining research attention since products and
services started to be exchanged on online platforms. Adomavicius et al. [
Approach: ActionRec Framework Components
Product text reviews contain a wealth of valuable information beyond the
numeric ratings expressed by buyers. Part of this information constitutes potential
actions that products enabled their buyers to perform. For example, a buyer
may articulate in her review that she is very happy with the tablet that she
bought, as it enabled her to work from home. Such expressed actions may ful ll
the needs of other buyers and drive their buying decisions. In marketing, such
buying decisions are often associated with the jobs-to-be-done concept, re ecting
that customers tend to buy products to help them do a certain job [
We propose ActionRec, a recommender system framework shown in Figure 1, to (1) capture the action-related entities in product text reviews through a semantic annotator ; (2) represent and link the entities in a knowledge graph; and (3) employ the knowledge graph data to build recommender apps.
Product Page
Buyer’s Preferences
Semantic Markup
Data Extractor Semantic Annotator
Product Data
Extractor Recommender App
Reviews Text Highlighter
Actions
Annotator Knowledge
Graph Recommender
Engine
Product Recommendation
Semantic Annotator. The semantic annotator component enables the annotation of product text reviews. It is currently implemented as a browser plugin to (1) automatically scan and extract relevant semantic markup data in product pages (e.g., description, o ers, pictures, etc. as per Schema.org1; (2) enable users to highlight text parts in the product reviews; and (3) annotate action-related elements (e.g., agent, environment, sentiment, etc.) The generated linked data is pushed to a triple-store hosting the knowledge graph.
Knowledge Graph. The knowledge graph is designed to connect the annotated text reviews to several entities including: Things such as Products, on which 1 https://schema.org/ potential Actions can be performed by Agents, who have Abilities to use certain product Features, in a certain Environment such as location, and the expressed Sentiments. To maximize the linking of data, we reused and extended entities de ned in existing vocabularies such as Schema.org and the Web Annotation Ontology2. The data is accessible online3.
Recommender Apps. The knowledge graph data is used to develop apps that provide product recommendations while integrating the captured actions in the process. At a high level, the apps take into consideration the buyer's speci ed preferences, combined with product data extracted from the knowledge graph through a SPARQL endpoint4. The extracted data is used by the recommender engine to display the recommended products and their related entities in the apps. Product recommendations can be provided by integrating product information about actions, review data, sentiments, and others, to assist online shoppers in their buying decisions. 3
We developed the ActionRec app, a web-based product recommendation prototype to demonstrate the potential integration of actions in the recommendation process5. Figure 2 shows a screenshot of the main features implemented.
The ActionRec app prototype relies purely on the knowledge graph data generated by the semantic annotator. So far, we have manually processed around 2,740 reviews of 9 computing devices from BestBuy6, with around 1,500 annotated actions. This generated around 58,500 triples in total. The prototype o ers several features including: - List of Products: The main window in the app (Part 1 of Figure 2) lists all the available products information available in the knowledge graph. - Most Relevant Actions: In addition to the products' feature details, the app displays the most prominent actions performed on the products (Part 2 of Figure 2). Such actions are detected by processing the knowledge graph data through a SPARQL query to extract the product, related actions, sentiment, and action count7. This data is used to compute the weighted ratio of positive sentiments out of the total expressed sentiments linked to this action. The weights of the actions are calculated based on the relative frequency of the actions performed on the product, extracted from the SPARQL query. For example, some of the most prominent actions performed on the Microsoft Surface Pro include \working," \studying," \carrying" and others. Clicking on the \more" link opens a tag cloud of actions. 2 https://www.w3.org/ns/oa 3 Product data page example: https://linked.aub.edu.lb/actionrec/Product/PUV 00001 4 SPARQL endpoint and sample queries page: https://linked.aub.edu.lb/actionrec/sparql 5 The prototype is accessible online: https://linked.aub.edu.lb/apps/actionrec/ 6 https://www.bestbuy.com/ 7 The query is available on the SPARQL endpoint page. 4 3 2 9 6 7
1 10 - Action-based Filter: On top of the regular feature-based lters shown in Part 3 of Figure 2, the app provides buyers the ability to nd products that ful ll certain needs expressed in the form of actions. In Part 4, users can specify their needs through the actions they are looking to perform, e.g., \work" and \draw," and the system will dynamically rank the products based on the products' relative weighted average positive frequencies of the entered actions. The user can specify the importance of the action using a slider. For example, in the current data, the Apple iPad Pro ranks higher compared to the Microsoft Surface Pro with respect to drawing and working with equal importance. However, putting higher importance on work pushes the Surface Pro on top. The current SPARQL query fetches products that support at least one of the actions entered by the user, and ranks the products accordingly. The query can be adapted to match the selection of products that support all the actions entered. - Detailed Actions View: The user can access a detailed view of the product by clicking on the \details" button in Part 5 of Figure 2. A new window will pop up, visualizing the actions and their frequencies in a bar chart (Part 6). The user can lter the actions based on the positive, negative, neutral sentiments (Part 7), or search for a speci c action (Part 8). The bars in the chart are clickable. For example, in Figure 2, the \draw" action is selected. Once clicked, the selected action's linked \agents," \environments," and \product features" are extracted from the knowledge graph and displayed in Part 9. At this level, the buyer can select speci c agents, environments, and features based on which the reviews and their annotations are ltered and displayed in Part 10. This way the reviews provide additional context to the annotated actions, and further assist the buyers in their purchase decisions in a more transparent and informed manner. 4
This paper demonstrated ActionRec, a novel approach for integrating the actions expressed in product reviews in the products recommendation process. The ActionRec framework includes a semantic annotator that helps with the automatic extraction of semantic markup data from product pages, enables users to highlight reviews, and annotate actions related to the products and other entities. The data is pushed to a publicly accessible knowledge graph, which can be used in recommender apps that process the action-related data to better match products with buyers' needs and recommend products accordingly.
This research can be extended in several directions. First, the action lters to specify user needs can be extended to allow capturing additional context elements such as the actions' objects (e.g., playing games versus playing music). Second, currently, the detection of actions is performed manually. This can be time-consuming and further work can be done on investigating the automation of actions and relations extraction from reviews. Third, a more detailed analysis of the approach would be needed to explore the potentials of the knowledge graph in performing certain recommendation tasks.
This work provides two contributions to the eld. First, it contributes to showing the potentials of o ering contextual product recommendations around the actions expressed in product reviews. Second, it provides an openly accessible knowledge graph that explicitly links user actions, products, and reviews context to computationally embed such data into the recommendation process.