Recognizing intentions of E-commerce consumers based on ant colony optimization simulation

Abstract

Identification of a consumer’s intent has a vital impact on commodity recommendation, selection of hot drainage commodity, website layout, and link settings. Most of the present studies on user intent are considered static. Specific intent is accompanied by a specific environment. Thus, intent is static when the environment does not change. However, the uncertainty of user access and purchase in e-commerce activities indicates that user intent can assume multiple forms and has multiple developmental stages. Therefore, this study draws support from the core ideas of an ant colony algorithm. Ants represent users, and pheromones represent user intent. User intents of browsing, collection, cart shopping, and purchasing behavior are obtained from ant responses to pheromones. Pheromone is expressed as the inner product of the objective attribute of commodity and user perception ability, because user intent is the matching result of objective attributes of commodity and subjective feelings of users, and its value is the concentration of user intent pheromone. Thus, the dynamics and uncertainty of user intention development can be presented by the ant colony algorithm. In this study, data were obtained from a NetLogo simulation experiment. We used neural networks to identify and verify user intentions of browsing, collection, cart shopping, and purchasing. The experimental results showed that the accuracy of intention prediction increased from 48% to 67%, and a level of the 11–20% accuracy improvement shows good, realistic predictions.

Keywords

Intent browsing purchasing pheromone ant colony optimization

1 Introduction

In electronic commerce activities, user intention is the psychological representation of understanding, attention, demand, and action before consuming behavior [1]. The meaning of user intention recognizes: (1) Commodity producers design and produce commodities that are in line with user intent [2];(2) E-commerce operators provide appropriate service strategies and present them to the consumers in a proper way that meets user intent [3]. Accurate identification of user intent is important for commodity recommendation, selection of hot drainage commodity, website layout, and link settings [4]. Therefore, accurate identification of user intent has become one of the most important areas of study in e-commerce. It attracts attention from a large number of researchers and has become the research focus in the field of e-commerce.

The identification of user intent is essentially a classification problem. The target structure in a classification problem has great influence on the classification method [5]. In accordance with the classification of target structure, user intent can be single structure, tree structure, and network structure, as shown in Fig. 1.

Fig.1

The structures of consumer intents.

As shown in Fig. 1, in the research of user intention recognition, scholars who hold the view of single layer structure assume that user intent is discrete and parallel to each other [6]. Therefore, what needs to be done is to list all of the candidate intents [7]. Then, a feature value is linked to a specific intent by a computer algorithm. This method is considered helpful for finding the rules of user intent. Scholars who hold the view of tree structure suggest that the idea of single layer structure is not enough to express the depth of user intent because it is a layer-by-layer depth, and subintent is the key point for expressing user intent. Therefore, identification of user intent needs the help of subintent identification. Compared with upper layer intent, each subintent describes more detailed and unique user needs. For this research method, identification of user intent is performed using tree structure to find the intent of each node. Scholars who hold the view of network structure suggest that intents are not separate but interrelated; therefore, intents can be transformed into each other under certain conditions. Most of the methods used by these scholars are based on graph theory.

The above methods have neglected the subjective initiative of user intention. Intention, environment, and the dynamic nature of conceptual work are separated, which leads to a result in which they can only find the user intent point but they cannot determine the intensity of user intent. This is the reason for low accuracy in user intent identification. The above intention recognition accuracy is generally around 55%. In most cases, it will be lower, which directly affects the operation efficiency and conversion rate of e-commerce.

In e-commerce activities, user intention has obvious uncertainties and dynamics [8]. Uncertainty reflects the fact that users are not aware of their own intentions and they are clear about their own specific target intention only under certain conditions [9]. Dynamic reflects the fact that the intensity of user intention is not the same under certain objective conditions: acquaintance, show interest, intention, even takes action [10]. The cause of uncertainty and dynamic of user intention has objective and subjective factors. Commodity prices, appearance, purchase amount, and other information are the objective factors. The user’s perception and sensitivity to the commodity appearance and purchase quantity are the subjective factors.

This study is different from existing methods such as knowledge base, bifurcation tree, support vector machine, and neural network. Considering the uncertainty and dynamic of intention, a simulation method based on an ant colony algorithm and a multi-agent system was adopted to cultivate the information data that represent consumer intention [11 –13]. The sources of forecast data were original intent structure data, commodity attribute data, and user perception ability data, combined with the cultivated information data. Neural networks were used to identify the transfer of suppliers (click on other commodity links), browse, collect, cart shopping, and purchase intention.

In this paper, the issue of user intention identification in e-commerce is discussed based on an ant colony algorithm. The second part analyzes the concept and basic attributes of user intention. The third part describes the basic content of an ant colony algorithm. The fourth part provides the mathematical expression and recognition process of user intent. In the fifth part, experimental results are shown, accompanied by the analysis. Finally, in the sixth part, we summarize this research study.

2 User intention

2.1 Concept

In e-commerce, user intention refers to a user’s mental representation of understanding, attention, needs, and actions before consuming behavior or behaviors related to consumption [14]. Intention is a causal relationship between desire and action, which is an initial manifestation of a previous behavioral goal [15]. In the actual e-commerce environment, user demand varies: user intention is reflected in the multilevel demand and the influence of different commodities at the same level [16]. Considering the influence of objective environment and subjective preference of the user, we can express the user intention as the vector in Equation (1). $P_{k, g} = I_{g} \cdot F_{k}$ (1)

In the formula, I_g is the objective attribute of commodity g; F_k is the consumer k’s subjective preference for commodities; P_k,g represents the intention of the consumer k for commodity g.

2.2 Uncertainty of user intent

User intent is uncertain; it varies under different circumstance. Therefore, it is situation-related. We can describe the e-commerce environment as the network diagram in Fig. 2.

Fig.2

A network representing e-commerce environments.

In Fig. 2, the network consists of node 0, node A, node B, node H, and connecting lines. Each node represents a commodity page; the lines connected to other nodes represent the hyperlink between relevant pages and the current page. After viewing the commodity page, which is represented by O, the next node that the user may turn to is determined by user intent. The next node may be B, C, D, E, A, or F. We describe them as nodes in a node set. The skip of intent is a probability event since intent is uncertain [17]. A browsing path is formed by the current node and the next node. Therefore, user intent will affect the choice of the user’s browsing path. In other words, the user’s browsing path also reflects the user’s intention.

2.3 Development of user intent

The purchase intention of a consumer has two prerequisites: (1) the product itself satisfies the consumer’s needs; (2) the consumer understands and trusts the product. From the perspective of understanding and trust, a consumer’s intention develops and has a life cycle. Commodity producers do a lot of advertising, sell products through the media to consumers or attract and forward services to increase understanding and trust, which also indicates that consumers’ intentions have life cycles [18].

Based on the AIDA marketing model proposed by marketing expert Heinz M. Goldmann, the life cycle of the intention of a consumer is divided into four stages, as shown in Fig. 3. The four stages are attention, interest, desire, and action. With the development of each stage, the intensity of consumption intention gradually increases [19].

Fig.3

Intent stages of e-commerce consumers.

In the attention stage of the intention development of consumers, the sites need to attract the attention of consumers through advertising and conspicuous image effects. In this stage, the behavior of consumers is not obvious until consumers are interested in the advertising or information. Then the intention would enter the stage of interest. Consumers in the interest stage will browse the information on products. If consumers think the product can bring satisfying benefits and they are able to pay, then the consumer intention enters the stage of desire. At the desire stage, the e-commerce runners should provide accurate descriptions of consumer benefits and product advantages compared with other similar products. In addition, the runners could inform consumers of the sales volumes of products to motivate them to purchase products. At this stage, if runners can provide promotional activities, consumers can easily enter the stage of action and pay online for the products.

From the viewpoint of the development process, in addition to subjective activities of consumers, the development of consumer intention also needs to be provided with relevant and suitable product information from e-commerce sites.

3 Intent identification

Consumers’ intentions are influenced by objective and subjective factors.

3.1 Commodity attributes

A number of scholars have conducted in depth studies on the influence of objective attributes of the commodity on the consumer’s intention. Through investigation of consumer online shopping behaviors, Zha and Wang found that consumers’ expectations and satisfaction of shopping services quality are the most important factors [20]. In the process of online shopping, the collection of goods, the number being reviewed, and the number being praised benefit the consumer in making the decision to buy goods. In contrast, the high price of goods and the sales amount of online shopping have a significant negative effect in consumers’ shopping decisions. In the B2C market, credit rating, consumers’ comments, and other credit-dependent factors are important variables that affect the purchasing preferences of consumers, and sales volume and price are other important variables that affect consumers’ preferences. By summarizing the objective factors discussed above, we select price, brand, comments, sales volume, and consumer preferences as the factors affecting the browsing and purchasing intentions of online consumers.

Considering the above analysis of consumer online shopping behaviors, we chose the following 5 aspects of the impact of online shopping: commodity prices (P), commodity brands (B), consumer comments (V), commodity sales volume (N), and consumers’ preference (L).

For different consumers, the objective values of the 5 factors of the same goods are the same. Among them, factors affecting the value of a commodity brand consist of brand awareness, brand value, and other brand-related factors. Consumer comments occur after consumers purchase and use goods. Factors that affect consumer comments include business service quality, timeliness of logistics, and user experience. Consumer comments play an important reference role in affecting consumer-purchasing behaviors. Commodity sales volume also plays an important role in affecting purchasing behaviors of consumers. In addition, consumer groups affect consumption behaviors of consumers. Thus, consumer group preferences will also affect consumer intention for purchasing.

3.2 Commodity attributes

The premise condition of the objective attribute of the commodity affecting consumer’s intention is that consumers have the ability to perceive. For example, if a consumer were more sensitive to the price of goods, the price of the goods would play a more important role in affecting the consumer’s purchasing behavior. If consumer f is sensitive to the price of the goods, then the subjective value k_pi is relatively large. In this way, the vector C_i (t) in Equation (2) can be used to present the subjective feeling of the consumer. $C_{i} (t) = (k_{pi}, k_{bi}, k_{vi}, k_{ni}, k_{li})$ (2)

In Equation (2), k_pi, k_bi, k_vi, k_ni and k_li are used to represent the perceptions of consumer i to the price, brand, consumer comments, commodity sales volume, and consumer preferences, respectively, at the moment t, and 0 ⩽ k_pi, k_bi, k_vi, k_ni, k_li ⩽ 1.

As shown in Equation (3), the vector O_j (t) represents the value of the attributes of the commodity j(j = 1, 2, 3, …, n) at the moment t. $O_{j} (t) = (p_{j}, b_{j}, v_{j}, n_{j}, l_{j})$ (3)

At the moment t, τ_ij (t) represents the actual influences of the goods j on the behaviors of the consumer i in Equation (4).

$\begin{matrix} τ_{ij} (t) & = & (C_{i} (t) \cdot O_{j} (t)) / η_{ij} \\ = & ((k_{pi}, k_{bi}, k_{vi}, k_{ni}, k_{li}) \cdot {(p_{j}, b_{j}, v_{j}, n_{j}, l_{j})}^{T}) / η_{ij} \\ = & (k_{pi} p_{j} + k_{bi} b_{j} + k_{vi} v_{j} + k_{ni} n_{j} + k_{li} l_{j}) / η_{ij} \end{matrix}$ (4)

In Equation (4), the greater the value of τ_ij (t), the more likely the consumer is to buy the goods.

As shown in Fig. 2, when the products are represented as commodity nodes, the links between the product pages and the recommendation links are formed as the transferring paths between nodes. Therefore, we can take the influence τ_ij (t) in Equation (4) as the pheromone in the ant colony algorithm.

3.3 The model of intent identification

Intention states are divided into 2 categories: intention development and intention transfer. When the consumer chooses to continue browsing and conduct other operations in the next operation, the consumer is in the intention development state. Otherwise, the consumer is in the intention transfer state.

3.3.1 Intention development

As intention recognition is performed in the process of interaction between consumers and e-commerce sites, to identify the intention of consumers in the process of interaction, we need to consider the consumer’s intention in the transfer and development process of consumer behavior. Figure 4 shows the different behaviors corresponding to different intention stages of consumers.

Fig.4

Behaviors of e-commerce consumers in different intent stages.

In the attention stage, e-commerce runners must attract the attention of consumers. If it succeeds, consumers will pay attention to hot commodity recommendation web pages, advertising banners, or search engines advertising information of e-commerce stores. In the interest intention stage, consumers will take the initiative to browse the shopping pages, compare the goods with other competing goods, or add the goods into the favorite collection for further operations. When consumers are in the desire intention stage, consumers will add the goods to the shopping cart. In the action stage, the consumer will place an order to achieve the intention.

To identify the transfer and development of consumer intention, we need to collect the data from consumer behaviors of browsing, collection, shopping cart, order, and purchase.

The intensity of the operation intention is calculated using Equation (4). Different intensity values of intention correspond to different operations. From low to high, according to the intensity value, the order of operation is to browse, to collect, to add to a shopping cart, or to purchase.

3.3.2 Transferring of intention

For consumer behaviors in e-commerce, the consumer intention includes the goods representing the direction of the consumer’s intention and the intensity value of the pheromone representing the active degree of the consumer’s intention. The consumer intention is represented as the form in Equation (5). ${v_{i} (t), C_{k} (t)} \to {v_{j} (t + 1), τ_{ij} (t + 1)}$ (5)

The constraints and the objectives of consumer intention recognition are given in Equation (8): in the moment t, the consumer k with the ability C_k (t) to perceive at the node v_i (t), select the next node v_j (t + 1), and the intensity of pheromone τ_ij (t + 1) representing the intensity of the intent. The probability of selecting v_j is given in Equation (6). $p_{ij}^{k} (t) = \frac{{[τ_{ij} (t)]}^{α} \cdot {[η_{jk} (t)]}^{β}}{\sum_{j} {[τ_{ij} (t)]}^{α} \cdot {[η_{ij} (t)]}^{β}}$ (6)

In contrast to the standard ant colony algorithm, the nodes and the paths that have been visited can be revisited. η_ij (t) can be computed using the formula in Equation (7). In the equation, i is the current node, and j is the candidate for the next node.

The intensity pheromone representing the intensity of the consumer’s intent is computed based on the formula in Equation (4). In this paper, the value of $Δ τ_{ij}^{k} (t)$ is 0 at the beginning. After that, the value of $Δ τ_{ij}^{k} (t)$ is determined by the consumers’ behaviors. When consumers choose a link or continue to stay on the webpage on behalf of a commodity to browse, collect, add to a shopping cart, and buy the product, the link or the webpage would be left with a certain amount of pheromone, and the pheromone on the paths or the goods would be increased. At this time $Δ τ_{ij}^{k} (t) > 0$ and the system is a positive feedback. If consumers do not browse the goods, collect the goods, add the goods to the shopping carts, or purchase the goods, then pheromone would not be increased, and $Δ τ_{ij}^{k} (t) = 0$ .

The probability of consumer k not to transfer to other goods but continue to browsing the current goods is given in equation (8). In the equation,η_ii = 1. $p_{i i}^{k} (t) = \frac{{[τ_{ii} (t)]}^{α}}{\sum_{j} {[τ_{ij} (t)]}^{α}}$ (8)

3.3.3 The algorithm of intent recognition

As shown in Fig. 5, the intention recognition algorithm is proposed as follows: (1) The network structure of the multi-agent based simulation is established according to the commodity attributes and link data of the network platform (network topology). At the same time, the perception ability of ants is set up. (2) Based on the data of the jump and other operations record of consumers, the program of the simulation is run. After running for a few ticks, the intensity of pheromone is convergent. In addition, the data of pheromone on the goods and paths are collected. (3) The information is used as the input of the neural network classifier or further simulation. The neural network is used to classify the consumers’ behaviors of browsing, collection, adding goods to the shopping carts, and purchase after training using historical data [21].

Fig.5

The algorithm of intent recognition.

4 Experiments and analysis

4.1 Experimental setup

In this research, we used NetLogo 5.0.4 of Uri Wilensky to conduct simulation experiments to verify the validity of the proposed model.

Considering the simulation system converges with 20,000 steps of simulation ticks, we set the ticks of the simulation to 20,000.

The system includes 2 types of pages: list pages and product pages. The list pages include the home page, the channel page, and promotion pages; product information is displayed on the product pages. The collection button, add-to-the-cart button, and other related product detail information may also show in the pages. $\begin{matrix} k_{pi} & = & \sum_{j = 1}^{J} k_{pij} \\ k_{bi} & = & \sum_{j = 1}^{J} k_{bij} \\ k_{vi} & = & \sum_{j = 1}^{J} k_{vij} \end{matrix}$ (9) $\begin{matrix} k_{ni} & = & \sum_{j = 1}^{J} k_{nij} \\ k_{li} & = & \sum_{j = 1}^{J} k_{lij} \end{matrix}$

The data of this experiment were from the early stage of an electronic commerce website. We recorded 21,619 access and other operation records of 183 consumers. After cleaning the data, we obtained 10,723 records of 100 consumers accessing 100 commodity pages as effective data for this experiment. The behaviors of browsing, collecting, adding to the shopping cart, purchasing, and clicking on the hyperlink to another good were recorded. Among those records, there were 7,271 records for browsing, collecting, adding to the shopping cart, and purchasing in the original pages, and 3,452 for clicking on the hyperlink to other goods.

Perception data of consumers were obtained by calculating the average good attribute values of the first 3–5 pages corresponding to each consumer. The calculating formula is given in Equation (9). In Equation (9), J = 3, 4, 5.

Tables 1 and 2 show data of the commodity attributes and consumer perception after being processed.

Table 1

Samples of good attributes

Attribute	Price	Brand	Consumer	Sales	Preference
			comments	volume
Good
g₁	0.43	1.00	1.00	1.00	1.00
g₂	0.51	1.00	0.50	0.45	1.00
g₃	1.00	0.50	0.00	0.09	1.00
g₄	0.51	0.50	0.50	0.27	0.50
... ..	... ..	... ..	... ..	... ..	... ..
g_i	0.46	1.00	0.50	0.10	0.50

Table 2

Samples of consumer perceptions

Perception	Price	Brand	Consumer	Sales	Preference
			comments	volume
Consumer
C₁	0.27	0.07	0.20	0.13	0.33
C₂	0.13	0.27	0.20	0.07	0.33
C₃	0.33	0.27	0.07	0.2	0.13
C₄	0.2	0.13	0.33	0.27	0.07
... ..	... ..	... ..	... ..	... ..	... ..
C_i	0.27	0.20	0.07	0.33	0.13

4.2 Experimental procedure

The steps of the experiment are as follows:

System initialization. In all, 100 commodity pages, 10 list pages, and 100 consumers were initialized. In the NetLogo interface, the product pages and the list pages are represented with circles, and consumers are represented with ants. The products are linked to each other using bidirectional links, forming a connected network. Figure 6 shows a screenshot of the simulation. At the same time, the adjustable variable parameters α, β and ρ were bound to the sliding controls on the NetLogo interface, so we could adjust the values by sliding the controls.

Parameter initialization. Set the value of the following variables: perception values k_pi, k_bi, k_vi, k_ni, k_li of m consumers (i = 1, 2, 3, … m), and the values p_j, b_j, v_j, n_j, l_j(j = 1, 2, 3, … n) of the commodity attributes. Pheromone intensity τ_ij (0) = 0 on the paths and products at the initial time.

Training. The values of the consumers’ perceptions in the simulation system were set according to the results obtained from the calculation using Equation (9) based on the data obtained from the real site. At the start of training, all behaviors of consumers were in accordance with the data collected, and the intensities of pheromone on paths and goods were updated using Equations (2) and (4).

Testing. Two testing methods were used to verify the validity of the training in Step (3): (1) Simulation testing. Consumers in the simulated system decide to stay on the current pages or click the links to other goods according to the probability given by the calculation of Equation (6); (2) Artificial neural network testing. Consumer behaviors of browsing, collection, shopping cart, purchasing, and clicking on links to other commodity pages were identified using a BP neural network [22]. The behaviors represent the intents.

The experiment ends.

Fig.6

A screenshot of the simulation.

4.3 Analysis and discussion of experimental data

4.3.1 Values of parameters α, β and ρ

Consumer’s intention is determined by intensity of the pheromone, so we set the values of parameters according influence on intensity of the pheromone. We conducted 1,331 experiments with 3 factors and 11 levels shown in Table 3 to investigate corresponding pheromones on paths and goods. The tick number of each experiment is 20,000.

Table 3
The factor levels

Levels

Factors 1 2 3 4 5 6

α 0.0 0.5 1.0 1.5 2.0 2.5

β 0.0 1.0 2.0 3.0 4.0 5.0

ρ 0.01 0.10 0.20 0.30 0.40 0.50

Levels

Factors 7 8 9 10 11

α 3.0 3.5 4.0 4.5 5.0

β 6.0 7.0 8.0 9.0 10.0

ρ 0.60 0.70 0.80 0.90 1.00

	Levels
α	0.0	0.5	1.0	1.5	2.0	2.5
β	0.0	1.0	2.0	3.0	4.0	5.0
ρ	0.01	0.10	0.20	0.30	0.40	0.50
	Levels
Factors	7	8	9	10	11
α	3.0	3.5	4.0	4.5	5.0
β	6.0	7.0	8.0	9.0	10.0
ρ	0.60	0.70	0.80	0.90	1.00

As shown in Table 4, we determined the values of 3 parameters based on the influences on mean and variance values of the pheromone in different factors with different levels.

Table 4

Correlation analysis between parameters and the intensity of the pheromone

	α
	Correlation	p-value	95% Confidence
	coefficient	p-value	interval
Average value	–0.0214	0.4332	[–0.0751 0.0322]
Variance value	0.0086	0.7529	[–0.0451 0.0623]
	β
Average value	–0.0068	0.8039	[–0.0605 0.0469]
Variance value	–0.0023	0.9317	[–0.0560 0.0513]
	ρ
Average value	0.0192	0.4839	[–0.0346 0.0728]
Variance value	–0.0131	0.6329	[–0.0667 0.0406]

Because all of the P values in Table 4 are far greater than 0.1, we accept the null hypothesis that the parameters α, β and ρ have no real effects on the distribution of the intensity of the pheromone.

Considering the convergence speed and computational efficiency of the ant colony algorithm, we set α = 1, β = 1 and ρ = 0.5 based on the suggestion of Ye and Zhang [23].

4.3.2 Influences of different consumer perceptions and behaviors on pheromone intensity

It can be concluded from the box diagram in Fig. 7 that the different consumer behaviors have a significant correlation with the pheromone intensity of commodities. In accordance with the order of the consumer behaviors of browsing, collecting, adding to the shopping cart, and purchasing, the values of intensity of the corresponding pheromone will gradually increase, which is consistent with the view of the AIDA model and this research. This further proves that we can use pheromone intensity based on consumer perception and the commodity attribute as an important index to identify the consumer’s intention.

Fig.7

Box diagram of pheromone intensities corresponding to different consumer behaviors.

4.3.3 Purchasing frequencies of consumers

Purchasing frequencies of consumers are shown in Fig. 8. Among the 93 consumers with repeated browsing habits before purchasing, 79 people would purchase again. The ratio of repeat customers to consumers who repeatedly browsed purchases was 84.95%. Among the repeat customers, the maximum number of purchases was 14 times.

Fig.8

Purchasing frequencies of consumers.

4.3.4 Recognition accuracy

Intentions were recognized based on the following factors: (1) the consumer’s perception abilities on price, brand, comments, sales volume, and public preference of goods; (2) goods attributes of price, brand, comments, volume and public preference; (3) pheromone intensity of the commodity; (4) the number of hyperlinks from the commodity to others; (5) the average value and the maximum value of pheromone intensities of hyperlinks.

We combined the above factors and employed the following 3 methods to recognize consumer intention: (1) A neural network was used to recognize the number of hyperlinks to other commodities from the current commodity being browsed, based on original data such as consumer perception abilities and commodity attributes; (2) A neural network was used to recognize intensities of pheromone obtained through NetLogo simulation, based on original data and intensities of pheromone on hyperlinks and commodities, and using steps (1), (2), and (3) given in part 4.2. Therefore, 7 factors were used to recognize intention. These factors consisted of consumer perception abilities, commodity attributes, the number of hyperlinks to other commodities, the average value of pheromone intensities on hyperlinks, and value of pheromone intensities on hyperlinks; (3) The method of NetLogo simulation was employed to recognize intention based on the same data used with method (2). Testing accuracy is given in Table 5.

Table 5
Accuracy of recognizing intentions

Factors Original Original data Original data

data + pheromone + pheromone

intensities from intensities from

the simulation the simulation

system system

Recognizing method Neural network Neural network Agent-based simulation

Browse (776 records) 43.30% 57.09% 64.82%

Collect (45 records) 55.56% 62.22% 71.11%

Add to the shopping cart (409 records) 56.47% 65.53% 76.04%

Purchase (205 records) 49.27% 51.22% 60.00%

Click hyperlinks to other goods (287 records) 51.57% 56.64% 65.85%

Total (1723 records) 48.81% 58.33% 67.21%

Factors	Original	Original data	Original data
Recognizing method	Neural network	Neural network	Agent-based simulation
Browse (776 records)	43.30%	57.09%	64.82%
Collect (45 records)	55.56%	62.22%	71.11%
Add to the shopping cart (409 records)	56.47%	65.53%	76.04%
Purchase (205 records)	49.27%	51.22%	60.00%
Click hyperlinks to other goods (287 records)	51.57%	56.64%	65.85%
Total (1723 records)	48.81%	58.33%	67.21%

In all, 10,723 records were accessed from the real system. In the experiment, we used 9,000 of the records as training data, and the remaining 1,723 records were used as testing data. Training data was used in 3 aspects: (1) training simulation system to obtain the intensities of pheromone on goods and paths of the system; (2) training the neural network for recognizing consumer intentions based on the original data; (3) training the neural network combined with the pheromone intensities from the simulated system.

The pattern recognition neural network (patternnet) of MathWorks MatLab 2011a was used to train and verify the proposed method (Goodman, Higgins, Miller, & Smyth, 1992). The numbers of hidden layers in methods (1) and (2) were 16 and 18, respectively, which were determined based on the method of “trial and error.”

From the results given in Table 5, we can see that the introduction of ant colony pheromone has a significant effect on improving the accuracy of e-commerce consumer intention recognition.

5 Conclusions

Consumer intention recognition has a vital impact on designing products sold through e-commerce channels and formulating business strategies for e-commerce operations. The difficulties of intention recognition lie in the dynamic and uncertain nature of consumer intention. Dynamic indicates that the intention is able to transfer and develop. Uncertainty indicates that there are probabilities for the transfer and development of the intentions. It is considered that the intention is mental representation before the occurrence of e-commerce behaviors, which is similar to pheromone intensity. Therefore, pheromone intensity is used to enhance the accuracy of recognizing intention of consumers in e-commerce.

Two difficult problems, dynamics and uncertainty in development of consumer intention, must be overcome to recognize consumer intention based on the ant colony algorithm: (1) For the problem of dynamics of consumer intention, the dynamics of pheromones in the ant colony algorithm can be used to represent the dynamics of consumer intention; (2) For the problem of uncertainty, the uncertainty in determining the path to visit next or staying at the current point can describe the uncertainty of consumer intention. In this research, the collected data from a real e-commerce website was used to train the agent-based simulation system to obtain pheromone intensities of various commodities and hyperlinks to improve the accuracy of recognizing consumer intentions [24].

The research results are as follows: (1) Introduction of pheromone from the simulation system based on the ant colony algorithm can greatly improve the accuracy of consumer intention recognition; (2) Setting values of parameters α, β and ρ of the ant colony algorithm has little impact on the distribution of pheromone intensities on commodity pages and hyperlinks; (3) Consumers who repeatedly browse connected commodity pages are likely to purchase and become repeat customers.

Footnotes

Acknowledgments

This research is supported by National Science Foundation of China (Grant No. 71203162), Science and Technology Planning Project of Guangdong Province, China (Grant No. 2014B040404072), Natural Science Foundation of Guangdong Province, China (Grant No. 2015A030313642), and Innovation Project of Wuyi University (Grant Nos. 2014KTSCX128 and 2015KTSCX144).

References

Armentano

M.G.

and Amandi

A.A.

, Personalized detection of user intentions, Knowledge-Based Systems24 (2011), 1169–1180.

Malhotra

, Galletta

D.F.

and Kirsch

L.J.

, How endogenous motivations influence user intentions: Beyond the dichotomy of extrinsic and intrinsic user motivations, Journal of Management Information Systems25 (2008), 267–299.

Barnes

, Hinton

and Mieczkowska

, The strategic management of operations in e-business, Production Planning and Control15 (2004), 484–494.

Gunasekaran

, Marri

H.B.

, McGaughey

R.E.

and Nebhwani

M.D.

, E-commerce and its impact on operations management, International Journal of Production Economics75 (2002), 185–197.

Guo

, Tang

and Liu

, Data association based on target signal classification information, Journal of Systems Engineering and Electronics19 (2008), 246–251.

, Zhou

, Hu

, Discovering and recognizing user’s intention based on pro-patterned extendable network in web active service, In: Proceedings of Fourth International Conference on Natural Computation, 2008, pp. 455–459.

Herlocker

J.L.

, Konstan

J.A.

, Terveen

L.G.

and Riedl

J.T.

, Evaluating collaborative filtering recommender systems, ACM Transactions on Information Systems22 (2004), 5–53.

Senecala

, Kalczynski

P.J.

and Nantel

, Consumers’ decision-making process and their online shopping behavior: A clickstream analysis, Journal of Business Research58 (2005), 1599–1608.

Hong

, Thong

J.Y.L.

and Tam

K.Y.

, The effects of information format and shopping task on consumers’ online shopping behavior: A cognitive fit perspective, Journal of Management Information Systems21 (2004), 149–184.

10.

Hsu

M.H.

, Yen

C.H.

and Chiu

C.M.

, and Chang, A longitudinal investigation of continued online shopping behavior: An extension of the theory of planned behor, International Journal Human-Computer Studies64 (2006), 889–904.

11.

Siebers

P.O.

, Maca

C.M.

, Garnett

, Buxton

and Pidd

, Discrete-event simulation is dead, long live agent-based simulation!, Journal of Simulation4 (2010), 204–210.

12.

Gotts

and Polhill

, Agent-based simulation in the study of social dilemmas, Artificial Intelligence Review19 (2003), 3–92.

13.

Zhang

and Zhang

, Agent-based simulation of consumer purchase decision-making and the decoy effect, Journal of Business Research60 (2007), 912–922.

14.

Kovac

V.B.

, Rise

, The role of desire in the prediction of intention, Swiss Journal of Psychology70 (2011), 141–148.

15.

Tan

and Ong

, A hybrid agent architecture integrating desire, intention and reinforcement learning, Expert Systems with Applications38 (2011), 8477–8487.

16.

and Wei

, Measuring the coverage and redundancy of information search services on e-commerce platforms, Electronic Commerce Research and Applications11 (2012), 560–569.

17.

Kuo

, Liao

and Tu

, Integration of ART2 neural network and genetic K-means algorithm for analyzing web browsing paths in electronic commerce, Decision Support Systems40 (2005), 355–374.

18.

Koufaris

, Kambil

and LaBarbera

P.A.

, Consumer behavior in Web-based commerce: An empirical study, International Journal of Electronic Commerce6 (2001), 115–138.

19.

Goel

, Hofman

J.M.

, Lahaie

, Pennock

D.M.

and Watts

D.J.

, Predicting consumer behavior with Web search, PNAS107 (2010), 17486–17490.

20.

Zha

and Wang

, Empirical study of influential elements of e-satisfaction, Management Sciences in China19 (2006), 50–58.

21.

Ozyildirim

B.M.

and Avci

, Generalized classifier neural network, Neural Networks39 (2013), 18–26.

22.

Liu

, Yan

and Pan

, Application of BP neutral network and genetic algorithm, The Open Cybernetics & Systemics Journal9 (2015), 1416–1421.

23.

and Zheng

, Configuration of parameters alpha, beta and rho in ant algorithm, Geomatics and Information Science of Wuhan University29 (2004), 597–601.

24.

Niazi

and Hussain

, Agent-based tools for modeling and simulation of self-organization in peer-to-peer, ad hoc, and other complex networks, IEEE Communications Magazine47 (2009), 166–173.