A Spreading Activation Model of Consumers' Asymmetric Similarity Judgment

Ehsan Ulhaque, University of Texas at Arlington
Kenneth D. Bahn, University of Texas at Arlington
ABSTRACT - Consumers often make asymmetric similarity judgments which are explained via Tversky's contrast model. This paper attempts to explain asymmetric similarity judgments by using the spreading activation models of semantic and declarative memory. It is proposed that the time taken for activation to spread from one node to the next is related to, and gives rise to, asymmetric similarity judgment.
[ to cite ]:
Ehsan Ulhaque and Kenneth D. Bahn (1992) ,"A Spreading Activation Model of Consumers' Asymmetric Similarity Judgment", in NA - Advances in Consumer Research Volume 19, eds. John F. Sherry, Jr. and Brian Sternthal, Provo, UT : Association for Consumer Research, Pages: 782-786.

Advances in Consumer Research Volume 19, 1992      Pages 782-786

A SPREADING ACTIVATION MODEL OF CONSUMERS' ASYMMETRIC SIMILARITY JUDGMENT

Ehsan Ulhaque, University of Texas at Arlington

Kenneth D. Bahn, University of Texas at Arlington

ABSTRACT -

Consumers often make asymmetric similarity judgments which are explained via Tversky's contrast model. This paper attempts to explain asymmetric similarity judgments by using the spreading activation models of semantic and declarative memory. It is proposed that the time taken for activation to spread from one node to the next is related to, and gives rise to, asymmetric similarity judgment.

INTRODUCTION

An essential component in the study of marketing and consumer behavior is that of consumer judgment and choice processes. These judgment and choice processes are based, to some extent, on cognitive representations -- consumers' internal memory structures -- as well as information which is acquired externally (Bettman 1979; Johnson and Puto 1987). While a plethora of research had been conducted on what, how and why of external information acquisition, research on cognitive representations of consumers has been rather limited (Gardner, Mitchell and Russo 1978; Johnson and Puto 1987). This state of affairs is rather sad as cognitive psychology provides a number of research streams and models in the area of cognitive representation (Anderson 1983; Collins and Loftus 1975; Smith, Shoben and Rips 1974). This paper provides an indirect testing of cognitive representation of brands in the tradition of Gardener, Mitchell and Russo (1978) and Meyers-Levy (1989).

A second stream of research in consumer judgment and choice process has dealt with the idea of similarity and proximity judgment. However, similarity relations have generally been dominated by geometric models in which objects are represented by points in a Euclidean space (Shepard 1974). Tversky (1977) has proposed a contrast model in order to overcome the metric assumption of geometric models. One of the strengths of contrast models is its capability to explain asymmetric similarity judgments. For example, according to geometric model of similarity judgment if a consumer feels that coke is similar to pepsi than it follows that she must feel pepsi to be equally similar to coke. However, consumers do make asymmetric similarity judgments. Thus, a consumer may find coke to be very similar to pepsi and at the same time pepsi to be less similar to coke (Johnson 1986). While Tversky (1977) has provided empirical support for the contrast model he has not related it to the various models of cognitive representation. This paper is an attempt in this direction.

Specifically, a network model of declarative memory (Anderson 1983) is proposed. The nodes of this network are occupied by brands with their related network of attributes. When nodes are activated, activation spreads to all other nodes connected to the activated node (Collins and Loftus 1975). The time taken for activation to spread from one node to another is hypothesized to be linked to similarity judgment. Asymmetric similarity judgment is proposed to be the result of differential time taken for the activation to spread from one node to another.

In the following sections we will first briefly describe the network model of memory. Then Tversky's contrast model of similarity judgment will be delineated. Finally, a conceptual model will be proposed to explain asymmetric similarity judgment via the network model of memory.

MODELS OF SEMANTIC MEMORY

Although various models of semantic memory have been proposed (e.g., Hierarchical-Network model of Collins and Quillian 1969; Feature-Comparison model of Smith, Shoben and Rips 1974; Spreading-Activation model of Collins and Loftus 1975) we will focus on spreading activation models. The reason being that the spreading activation, or network, models have better fared the scholarly scrutiny of various researchers (Chang 1986).

According to the spreading activation model of Collins and Loftus (1975), the concepts (or brands in this case) are represented in memory as nodes, and relations between brands are represented as associated pathways between the nodes. There could be multiple paths of various strengths and lengths between brands. For example, coke could be one such node in a consumer's memory while different attributes of coke (carbonated, cola, distinctive taste, etc.,) could be connected to the node 'Coke' by associated pathways. Another brand of soft drinks (e.g., pepsi) may share some of the attributes of coke (e.g., carbonated) while it may also have some unique attributes of its own. The attribute 'carbonated' is thus linked both to coke and pepsi. The more attributes two brands have in common, the more links there are between the two nodes, and the more related in memory the two brands will generally be. When a brand is activated, for example when attention is focused on coke, the activation spreads out in a decreasing gradient along the links emanating from the activated brand. This activation may reach the node 'Pepsi' via the attribute node 'carbonated'. When a path is found between two brands, i.e., two spreading activations intersect, a decision regarding them could be made (Collins and Loftus 1975). Cognitive psychologists seem to use the concepts of activation spreading from node 'a' and reaching node 'b', and spreading activations from node 'a' and node 'b' intersecting interchangeably (Ratcliff and McKoon 1981).

Anderson (1983) has proposed a similar process in his ACT* model of human memory. According to his model, brands or nodes have an activation value associated with them which reflect the current importance, relevance, or availability of the node in memory. The links between nodes have weights associated with them which are determined by the strength of associations between the connected nodes. For example, if coke is regarded as the most popular soft drink then the link associating node 'Coke' to node 'soft drink' will have more weight associated with it reflecting the strength of association between the two nodes. On the other hand, if coke is not thought of as a food item then the weight of the link associating node 'Coke' to node 'food' will be relatively low, or there will be no such link. When a node is activated the spread of activation is modulated by the weights associated with each link, i.e., closely associated nodes produce higher activation transfer as compared to weakly associated nodes. Reder and Anderson (1980) found that the amount of activation spreading from a given node along a pathway is a function of the strength of that pathway relative to the sum of strengths of all paths emanating from that node.

Priming tasks have provided sufficient empirical support to network models of memory (Ashcraft 1989). In priming tasks a prime stimulus, which is presented first, influences some later process on a target stimulus which follows the prime. For example, in a lexical decision task (a task in which subjects have to respond whether a string of letters form a regular word or not) it was found that subjects took less time to recognize the target 'butter' as a regular word when primed by the word 'bread' than the time they took to recognize 'butter' as a word when primed by the word 'doctor' (Meyer and Schvaneveldt 1971). The spreading activation explanation of this differential response speed is as follows. The prime stimulus 'bread' activates the node 'bread' in the subject's memory. As 'bread' and 'butter' are related concepts, the link associated with the two nodes of bread and butter have a relatively higher weight associated with it. The activation spreads from 'bread' towards 'butter'. When the subject is asked whether butter is an English word the decision is facilitated as the node 'butter' is already activated because of the spread of activation from the prime 'bread'. On the other hand, as doctor and butter are not related concepts, therefore, when doctor is used as prime, activation does not spread towards butter, and consequently the lexical decision is not facilitated.

An important concept in Anderson's model is the 'fan effect' which means that performance of subjects on various tasks (e.g., lexical decision tasks as explained above) will slow down on any fact if that fact is but one of many that have been learned about a particular concept. The rationale is as follows. Given a fixed level of activation for a node the more links it has, the less activation is available for any one link. For example, if the node 'bread' is only linked to node 'butter' then all of the available activation of 'bread', when it is primed, will go towards 'butter'. Thus, the lexical decision for butter will be facilitated. However, if 'bread' is connected to 'breakfast' as well as 'butter' and 'sandwich', then only one third of 'bread's' activation can reach 'butter' as the activation will spread in all three links. Consequently, the lexical decision for butter will not be as facilitated as when only one link was present. In a consumer behavior context, Meyers-Levy (1989a) have shown that brand memory for high frequency brand names is adversely influenced if their association sets are large - a phenomenon similar to the fan effect.

The concept of spreading activation has been used to explain mechanisms involved in such tasks as category exemplar production (Loftus 1973), semantic priming in lexical decisions (Neely 1977), episodic sentence and word recognition (Anderson 1983), and reading comprehension (Stanovich and West 1983). It has been experimentally found that the spread of activation is automatic as opposed to being under strategic control when the time duration between the prime and the target is less than 250 milliseconds (Neely 1977). Conscious processes, however, may take over at longer durations (Posner and Snyder 1975). It has also been found that the activation spreads beyond directly related concepts, although this 'mediated priming' effect is only evidenced in a pronunciation task (Balota and Lorch 1986). Although not in the context of spreading activation, priming paradigm has also been used in the area of consumer behavior. Thus, Herr (1989), Meyers-Levy (1989b) and Yi (1990) have documented the impact of priming on consumer information processing and judgment. This study will, however, use priming in order to model similarity judgment tasks.

CONTRAST MODEL OF SIMILARITY JUDGMENT

The theoretical analysis of similarity relations in marketing has been dominated by geometric models in which products, or brands, are represented by points in some coordinate space where the metric distance between the points reflect the similarities between respective brands (Green and Rao 1972). One result of such a dimensional representation is that the similarity between brand 'a' and 'b' (or Euclidean distance between them) remains the same whether one measures similarity of 'a' to 'b' or of 'b' to 'a'. Tversky (1977) has, however, argued that the metric assumption that underlie the geometric approach to similarity are often found lacking. He has developed an alternative feature-based approach to the analysis of similarity relations. According to his 'contrast' model, each brand is characterized by a set of features or attributes. For example, coke has attributes of being a carbonated drink, a cola drink, among others. The observed similarity between two brands is expressed as a function of their common and distinctive attributes. Thus, if brand 'a' (e.g., coca-cola) has a set of attributes 'A' and brand 'b' (e.g., shasta cola) has a set of attributes 'B' then the similarity between them is expressed as:

S (a,b) = q f(A 1 B) - a f(A - B) - b f(B - A)

The model allows for a variety of similarity relations over the same set of objects through its flexibility of designating different values for q, a, b as well as different functions f. Thus, in contrast to the dimensional model, similarity of 'a' to 'b' can be different from similarity of 'b' to 'a'. The factors that contribute to the salience of a stimulus, i.e., intensity, frequency, familiarity, good form, informational content, etc., are reflected by the scale 'f' in the contrast model. Variations in similarity judgments because of the 'task' or 'context' effects are captured by the parameters q, a, and b.

FIGURE 1

A MODEL OF MEMORY NETWORK

Tversky and his colleagues have published a few papers lending support to the contrast model. For example, Tversky and Gati (1978) have used the contrast model to explain asymmetric similarity judgments. They have shown that subjects' judged similarity of North Korea to China exceeds their judged similarity of China to North Korea. In a consumption context, Johnson (1986) has found that consumers rated the similarity of shasta cola to coca cola to be higher than the similarity of coca cola to shasta cola. Thus, coke is not as similar to shasta as shasta is to coke. Tversky and Gati (1978) explain this asymmetry via a focusing hypothesis and a relative salience effect in their contrast model. The focusing hypothesis states that if S(a,b) is interpreted as the degree to which 'a' (subject) is similar to 'b' (referent) then in such judgments attributes of 'a' are weighted more heavily than attributes of 'b' (i.e., a > b). The salience effect follows from the contrast model. With a > b,

S(a,b) > S(b,a)                 iff

f(B-A) > f(A-B)

or when the distinctive attributes of 'b' are more salient than the distinctive attributes of 'a'.

While the explanation provided by Tversky and Gati (1978) is reasonable, however, it is not grounded in the memory network models. For example, how are 'subject' and 'referent' cognitivly represented and how features or attributes are given more weight than less. In the following section a memory network model of asymmetric similarity judgment is presented.

A SPREADING ACTIVATION MODEL OF ASYMMETRIC SIMILARITY JUDGMENT

Assume that 'a' (e.g., coke) is a node in the semantic memory. This node is linked to many attributes like carbonated, refreshing, dark color, sweet taste, distinctive bottle, etc., (Gardener, Mitchell and Russo 1978). If A is the set of these attributes of coke then node 'Coke' has A different links emanating from it. Similarly, node 'b' (e.g., shasta) has B different links emanating from it. Now some of the attributes (A*B) are common to coke and shasta (e.g., carbonated, dark color), while some attributes (A-B or B-A) are distinct to either coke or shasta (e.g., distinctive bottle, world-wide sales). Thus, the network structure might be represented as in Figure 1.

The focusing hypothesis of Tversky and Gati (1978) is proposed to be captured by a priming effect. According to Tversky and Gati when coke is the subject and shasta the referent, then the attributes of coke are weighted more heavily than the attributes of shasta. However, they also suggest that human subjects have a propensity to use directional statements such as "x is like y" where the subject 'x' always comes before the referent 'y' in similarity statements. A case can be made that the subject 'x' is acting as a prime for the referent 'y'. Thus, in a similarity estimation of coke to shasta, it is proposed that coke acts as a prime for shasta. 'Coke' becomes a source of spreading activation which reaches 'Shasta' via the common attribute links.

The salience of a stimulus depends on factors which include intensity, familiarity, frequency, informational content among others (Tversky and Gati 1978). All of these factors may safely be assumed to give rise to two phenomenon. Firstly, they will increase the steady state level of activation of a node (Anderson 1983). Secondly, they will increase the number of links emanating from a node as well as their strengths (Collins and Loftus 1975). Thus in the previous diagram, coke is more salient than shasta (coke has more links than shasta). Because of the 'fan effect', however, 'Coke' will be able to spread a lesser amount of activation to 'Shasta' via the common links as compared to that sent by 'Shasta' to 'Coke', assuming that the two have an equal amount of steady state activation level.

The process of asymmetric similarity judgments between coke and shasta can now be grounded into a proposed semantic network model. When a subject is judging the degree to which shasta is similar to coke, node 'Shasta' becomes the focused unit. Activation spreads from 'Shasta' via the mediating common attributes to 'Coke' (Balota and Lorch 1986). The strength of activation reaching 'Coke' depends on the sum of weights of the common links between 'Shasta' and 'Coke' relative to the weights of all links emanating from 'Shasta'. The speed with which the intersection of activations from 'Shasta' and 'Coke' takes place, and the information about the level of activation (which is available to the subject; Anderson 1983) guides the subject to decide on the similarity of shasta to coke. A similar process occurs when the subject is determining the degree to which coke is similar to shasta. If we assume that the steady state activation level of shasta and coke are the same and also that coke is more salient than shasta, then because of the fan effect less activation will reach shasta from coke, as coke has many more links which also have greater strength, while more activation will reach coke from shasta. Because of this difference in activation transfer, the speed with which a critical level of activation intersection is reached will be more in the case when shasta is the prime than when coke is the prime (Anderson 1983). The information of this level of activation at the intersection is proposed to result in a degree of similarity judgment. According to the 'semantic relatedness effect' those concepts that are highly interrelated can be judged more rapidly in a lexical task than those with lower degrees of relatedness (Ashcraft 1989). Thus, speed of spreading activation and relatedness (or similarity in this case) are related. The more the activation spreads, the sooner the intersection, and consequently the more similar the prime concept is to the target concept. Thus, shasta will be reported as more similar to coke than vice versa. This asymmetric similarity judgment is exactly the same as that obtained by Johnson (1986). However, the proposed model presented in this paper provides asymmetric similarity judgment a grounding in memory network models. Based on the above logic we would hypothesize the following:

A more salient brand within a product category, when used as a prime, will slow the lexical decision or pronunciation task for a less salient target brand within the same product category. Thus, the more salient brand will be judged as 'less' similar to less salient brand.

On the other hand, a less salient brand within a product category, when used as a prime, will facilitate the lexical decision or pronunciation task for a more salient target brand within the same product category. Thus, the less salient brand will be judged as 'more' similar to more salient brand.

CONCLUSIONS

A spreading activation model of consumer's asymmetric similarity judgment is proposed. The model, if validated by empirical data, could lead to many interesting implications in the area of consumer's memory of brands and choice process. Firstly, it directly impacts on the contents of a consumer's evoked set. It seems that because of differential activation capabilities of less salient and more salient brands the composition of evoked set will depend on the 'cue' brand. If the cue brand is less salient then the probability of recall of more salient brand will be high and so will be its probability of being a member of the evoked set. On the other hand, if the cue brand is more salient then the probability of recall, and hence membership in evoked set, of a less salient brand will be low.

Secondly, if according to this model priming has an important role to play in similarity judgment then an area of automatic and controlled similarity judgment might evolve. Posner and Snyder (1975) had shown that automatic priming precedes controlled processes. Thus, one may argue that in a low involvement choice process, where controlled processes are not dominant, automatic activation of brands may determine the similarity judgments among brands. This similarity judgment will in turn influence the choice. On the other hand, in a high involvement choice process, where controlled processes dominate, the automatic activation of brands will not determine the similarity judgments among brands and, consequently, the choice may turn out to be very different. Finally, this automatic or controlled similarity judgment may also impact brand substitution and brand switching behavior.

REFERENCES

Anderson,J.R. (1983)," A Spreading Activation Theory of Memory," Journal of Verbal Learning and Verbal Behavior, 22, 261-295.

Ashcraft, M.H. (1989), Human Memory and Cognition, Illinois: Scott, Foresman and Company.

Balota, D.A., and Lorch, R.F.(1986), "Depth of Automatic Spreading Activation: Mediated Priming Effects in Pronunciation but not in Lexical Decisions," Journal of Experimental Psychology: Learning , Memory, and Cognition, 12, 336-345.

Bettman, J.R. (1979), An Information Processing Theory of Consumer Choice. Reading, MA: Addison-Wesley Publishing Company.

Chang, T.M. (1986), "Semantic Memory: Facts and Models," Psychological Bulletin, 99, 199-220.

Collins, A.M., and Loftus, E.F. (1975), "A Spreading Activation Theory of Semantic Processing," Psychological Review, 82, 407-428.

Collins, A.M., and Quillian, M.R. (1969), "Retrieval Time From Semantic Memory," Journal of Verbal Learning and Verbal Behavior, 8, 240-247.

Gardner, M.P., Andrew, A.M., and Edward R. (1978), "Chronometric Analysis: An Introduction and an Application to Low Involvement Perception of Advertising," Advances in Consumer Research, Vol 5, H. K. Hunt, ed. Ann Arbor, MI: Association for Consumer Research, 581-9.

Green, P.E. and V.R. Rao (1972), Applied Multidimensional Scaling: A Comparison of Approaches and Algorithms. New York: Holt, Rinehart and Winston, Inc.

Herr, Paul M. (1989), "Priming Price: Prior Knowledge and Context Effects," Journal of Consumer Research, 16 (June), 67-75.

Johnson, M.D. (1986), "Consumer Similarity Judgments: A Test of the Contrast Model," Psychology and Marketing, 3 (1), 47-60.

Johnson, M.D. and C.P. Puto (1987), "A Review of Consumer Judgement and Choice," Review of Marketing, 1987, ed. M.J. Houston. Chicago: American Marketing Association. 236-292.

Loftus, E.F. (1973), "Activation of Semantic Memory," American Journal of Psychology, 86, 331-337.

Meyer, D.E., and Schvaneveldt, R.W. (1975), "Meaning, Memory Structure, and Mental Processes," The Structure of Human Memory, ed. C.N. Coffer. San Francisco: Freeman. 54-89.

Meyers-Levy, Joan (1989a), "The Influence of a Brand Name's Association Set Size and Word Frequency on Brand Memory," Journal of Consumer Research, 16 (Sept), 197-207.

Meyers-Levy, Joan (1989b), "Priming Effects on Product Judgments: A Hemispheric Interpretation," Journal of Consumer Research, 16 (June), 76-86.

Neely, J.H. (1977), "Semantic Priming and Retrieval from Lexical Memory: Roles of Inhibitionless Spreading Activation and Limited Capacity Attention," Journal of Experimental Psychology: General, 106, 226-254.

Posner, M.I., and Snyder, C.R.R. (1975), "Attention and Cognitive Control," Information Processing and Cognition, ed. R.L. Solso. Hillsdale, NJ: Erlbaum.

Ratcliff, Roger and Gail McKoon (1981), "Does Activation Really Spread?" Psychological Review, 88, No 5, 454-462.

Reder, L.M., and Anderson, J.R. (1980), "A Partial Resolution of the Paradox of Interference: The Role of Integrating Knowledge," Cognitive Psychology, 12, 447-472.

Shepard, R.N. (1974), "Representation of Structure in Similarity Data: Problems and Prospects," Psychometrika, 39, 373-421.

Smith, E.E., Shoben, E.J., and Rips, L.J. (1974), "Structure and Process in Semantic Memory: A Featural Model for Semantic Decisions," Psychological Review, 81, 214-241.

Stanovich, K.E., and West, R.F. (1983), "On Priming by a Sentence Context," Journal of Experimental Psychology: General, 112, 1-36.

Tversky, A. (1977), "Features of Similarity," Psychological Review, 84, 327-352.

Tversky, A. and Gati, I. (1978), "Studies of Similarity," Cognition and Categorization, ed. E. Rosch and B. Lloyed. Hillsdale, NJ: Erlbaum. 79-98.

Yi, Youjae (1990), "The Effects of Contextual Priming in Print Advertisements," Journal of Consumer Research, 17 (September), 215-222.

----------------------------------------