Utilizing multi-source data in popularity prediction for shop-type recommendation

Highlights

• Multi-source information and big data analytics are utilized to study this problem.

• Location profiles are constructed considering internal and external characteristics.

• The location profile and the commercial structure are integrated.

• The principle of CF is introduced to deal with shop-type recommendation problem.

• The performances of the proposed method are validated on a real-world dataset.

Abstract

It is important for an investor to determine the most suitable shop type (e.g., restaurant, cafe) given a location. Traditionally, investors determine shop types based on their subjective judgments and perceptions. However, insufficient information and cognitive limitation often lead to flawed decisions and increase investment risks. With advances in information technology, multi-source heterogeneous information and big data analytics can be utilized to provide support for making such a decision. In this paper, we propose a novel shop-type recommendation method that suggests a suitable shop type based on multi-source information collected from a business review site, a location-based navigation system and a mobile carrier. Specifically, our method first constructs the location-type matrix to alleviate the problem of incomplete data and models the location profile by considering internal and external features simultaneously. In particular, a hybrid similarity model is proposed to integrate the location profile and the commercial structure into a unified framework. Then, a location-based collaborative filtering method is developed to predict shop popularity and suggest a suitable shop type. Finally, we demonstrate the effectiveness of our method compared to several benchmark methods by applying it to a real-world dataset from China.

Introduction

It is an important but challenging task for an investor to determine the most appropriate shop type for a given location [1]. Shop-type recommendation aims to suggest an optimal shop type (such as restaurant, cafe, and bar) in a given commercial location to attract more consumers. Selecting a suitable shop type is not only beneficial to reducing investment risks, but also crucial for the development of community economy and future business [1].

One of the key problems for shop-type recommendation is predicting shop popularity which can be represented by the number of attracted consumers [1], [2]. Features, which are closely related to shop popularity, can be categorized into two classes: the external characteristics (e.g., geographical environment [3], people flow [4] and commercial structure [5]) and the internal characteristics (e.g., environment [6], service [7] and product quality [8]). The impacts of these features vary on different shop types. For example, accessibility is more important for a convenience store while good service is more important for a cafe. In the real-world, investors usually make decisions based on their subjective judgments and perceptions on these features. Due to insufficient information and the limitation of cognition, it is difficult for investors to quantify and distinguish the impacts of multiple features. This may lead to flawed decisions and thus increase the subsequent investment risks. Therefore, it is important to collect sufficient information concerning these features and propose an effective approach to determine the suitable shop type for investors.

During the last decade, the rapid development of location-based services and big data mining technologies have made it possible to collect and analyze multi-source data from various platforms for decision support. Business review sites, such as Dianping.com,¹ contain shop-related information and customer ratings. Location-based navigation systems, such as Ditu.amap.com,² provide information on numerous points of interest (POI) and their relevant information. Moreover, mobile carriers collect pervasive spatio-temporal trajectories of people’s daily movement. With these data sources, the aforementioned shops’ external and internal characteristics can be evaluated. In this paper, we utilize these data to support the recommendation. One example of the data is depicted in Fig. 1. However, after obtaining such data, there is still a challenge during the recommendation: how to provide decision support of shop-type recommendation by integrating multi-source heterogeneous data.

In order to address this challenge, we propose a shop-type recommendation method integrating multi-source information to predict shop popularity for shop-type recommendation. First, we collect multi-source data from business review sites, location-based navigation systems and mobile carriers and preprocess the data using the MapReduce framework implemented with the Hadoop platform. Subsequently, we construct a location-type matrix by aggregating shops’ popularity to alleviate the problem of incomplete data. Then, the location profile and the commercial structure are constructed from multi-source data to describe the location’s features at micro and macro levels. In addition, a hybrid similarity measurement is proposed to measure the closeness between locations with the consideration of the location profile and the commercial structure. Finally, on the basis of the location-type matrix and the hybrid similarity, a location-based collaborative filtering is adopted to predict shop popularities for all shop types. For a given shop location, the top-$n$ shop types with the highest predicted popularities are recommended to investors.

The main contributions of our work can be summarized as follows.

First, multi-source information and big data analytics are utilized to study the shop-type recommendation problem. In comparison with traditional ways of decision making based on investors’ subjective judgments and perceptions, the proposed method considers more relevant factors and provides reliable decision support for investors to overcome cognitive limitation and reduce investment risks.

Second, location profiles are constructed from multiple aspects by considering internal and external characteristics simultaneously, which provides a comprehensive description of locations. Particularly, internal characteristics are analyzed in this problem for the first time and help to characterize location profiles more accurately.

Third, a hybrid similarity model integrates the location profile and the commercial structure into a unified framework. The location profile and the commercial structure are used to measure the similarity between locations at micro and macro levels. The combination of the two perspectives contributes to improving recommendation performance.

Fourth, the principle of collaborative filtering is introduced to deal with the shop-type recommendation problem for the first time, where a location-type matrix is proposed to alleviate the problem of data incompleteness. The proposed method takes full advantage of commercial information of other locations to provide valuable references for recommending a suitable shop type to the target location.

Finally, the performances of the proposed method are validated through comparison with several benchmark methods on a real-world dataset. This dataset is collected from a business review site, a location-based navigation system and a mobile carrier, related to 29673 shops, 10965 locations and 17 million individuals in Beijing, China.

The remainder of this paper is organized as follows. Section 2 reviews the background and related works. Section 3 describes the problem and introduces the research framework. Section 4 describes the process of recommendation generation in detail. The experiments and results analysis are discussed in Section 5. Finally, Section 6 gives a conclusion including our contribution and the direction of future work.