Schemata in Consumer Research: a Connectionist Approach

Tom J. Brown, University of Wisconsin-Madison
ABSTRACT - Schemata and related knowledge structures have figured prominently in consumer research in recent years, yet little consideration has been given to the processes through which schemata are retrieved from memory. In cognitive psychology, schemata have typically been considered frameworks of knowledge that are stored in memory and retrieved through a process of spreading activation. From a connectionist perspective, however, schemata are not stored at all. Rather, they are re-created as individuals need them; schemata are simply patterns of activation across a network of units connected in a constraint satisfaction system. A connectionist model of schemata is presented and demonstrated in a consumer context.
[ to cite ]:
Tom J. Brown (1992) ,"Schemata in Consumer Research: a Connectionist Approach", in NA - Advances in Consumer Research Volume 19, eds. John F. Sherry, Jr. and Brian Sternthal, Provo, UT : Association for Consumer Research, Pages: 787-794.

Advances in Consumer Research Volume 19, 1992      Pages 787-794

SCHEMATA IN CONSUMER RESEARCH: A CONNECTIONIST APPROACH

Tom J. Brown, University of Wisconsin-Madison

[The author expresses his appreciation to Professors Arthur M. Glenberg and Michael L. Rothschild of the University of Wisconsin-Madison.]

ABSTRACT -

Schemata and related knowledge structures have figured prominently in consumer research in recent years, yet little consideration has been given to the processes through which schemata are retrieved from memory. In cognitive psychology, schemata have typically been considered frameworks of knowledge that are stored in memory and retrieved through a process of spreading activation. From a connectionist perspective, however, schemata are not stored at all. Rather, they are re-created as individuals need them; schemata are simply patterns of activation across a network of units connected in a constraint satisfaction system. A connectionist model of schemata is presented and demonstrated in a consumer context.

INTRODUCTION

Considerable research and theory in consumer behavior has been based on the notions of categories or schemata possessed by consumers (e.g., Cohen and Basu 1987; John and Sujan 1990; Loken and Ward 1990; Meyers-Levy and Tybout 1989; Sujan and Bettman 1989; Sujan et al 1986). According to the standard theory in cognitive psychology, schemata are stored frameworks of knowledge about some object or topic and are represented by nodes in semantic memory. Schemata are retrieved into working memory through the process of spreading activation (Ashcraft 1989).

Many cognitive phenomena, including schema development and retrieval, have been addressed in recent years from a different perspective, alternatively referred to as connectionism, parallel distributed processing (PDP) models, or neural networks. Such approaches, which feature formal mathematical models of cognitive processes, have enjoyed increasing popularity (Hintzman 1990). In contrast to the standard theory, connectionist approaches to schemata assume that schemata are implicit and created at the time an individual needs them. In short, schemata are not "things" stored in memory; rather, they are simply patterns of activation across a vast network of units. The connectionist approach offers a conceptually distinct model of schemata that approximates the physiology of the brain while demonstrating many of the recognized properties of schemata.

In this paper, one particular connectionist model of schemata retrieval based on a constraint satisfaction system is presented and applied in a consumer context. Data were collected with respect to a given product class (automobiles) in order to demonstrate the model.

A CONNECTIONIST APPROACH TO SCHEMATA

In general, connectionist models "assume that information processing takes place through the interactions of a large number of simple processing elements called units, each sending excitatory and inhibitory signals to other units" (McClelland et al 1986, pg. 10). These models are distinguished from traditional theories of learning and memory in several important ways. In general, they are more closely tied to the physiology of the brain. The brain consists of a large number of interconnected elements which appear to send inhibitory or excitatory impulses to other elements; moreover, these elements appear to update their excitations on the basis of these simple impulses (McClelland et al 1986).

In connectionist models, storage in memory does not involve the actual storage of semantic or episodic information in the form of nodes. Instead, the connection strengths or weights between units are stored, allowing information to be re-created at the time an individual needs it. Most connectionist models of information processing do not require the presence of a superordinate control function for organizing and storing information or guiding retrieval. Rather, the system learns of its own accord the connection strengths between units through the repeated processing of stimuli over time. Importantly, multiple patterns may be represented by the same set of units since it is not the units themselves that are important, but rather the connection strengths, or weights, between units.

The particular connectionist model presented in this paper is a constraint satisfaction model. The operation of the constraint satisfaction network in re-creating a schema provides perhaps the greatest point of contrast with the standard theory of schema retrieval. In a spreading activation model, activation moves from one activated node to other related nodes via connecting pathways (with the nodes themselves representing semantic memory). As the activation spreads from node to node, the schema is retrieved. In a connectionist constraint satisfaction model, all units are connected to all other units, and an activated unit may inhibit some units while exciting others. All the units may thus play a role in determining the final pattern of activated units. Moreover, this final pattern of activation across units may not be accounted for by a spreading activation model because units that are activated at some point may be inhibited (and ultimately de-activated) by other units in the network as constraints are satisfied.

The units in a constraint satisfaction model may be thought of as hypotheses about the presence of certain features or attributes. In a consumer marketing context, consider the schemata that individuals hold with respect to different types of automobiles. One unit might represent a hypothesis about the presence of high status or prestige. (In reality, the highly abstract attributes used in this example are themselves likely to be patterns of activation across a great number of lower-level units.) This unit may be activated as part of a schema for certain types of automobiles (e.g., expensive sports cars), but not activated as part of a schema for other types (e.g., economy cars). The connection between two units represents a constraint between those units. If the units tend to co-occur (e.g., units representing high status and high price), then a positive weight is given to the connection between the units. If one unit tends to be activated while the other unit is not (e.g., units representing high status and vinyl seats) then a negative weight is "learned" by the system for the connection between the units. In some situations, there may be little regularity in the co-occurrence of units, leading to a connection weight approaching zero (e.g., units representing an airbag and leather interior).

By expanding the model to include multiple units, the workings of the constraint satisfaction network can be seen. Each unit receives excitatory inputs from units to which it is connected with positive connection weights, and inhibitory inputs from units to which it is connected with negative connection weights. The system will settle, or "relax", into a stable state where as many constraints as possible are satisfied, with priority given to those constraints (connections) weighted most heavily. In this model, the units that are activated in the stable state represent the relevant schema.

In addition to receiving inputs from other units, a unit may also receive input externally. This external input, or activation, represents external evidence that the feature represented by a unit is present, and may serve to set the system in motion. For example, the viewing of a print advertisement for the Ford Escort may be sufficient to re-create the full economy car schema for many individuals. In the following demonstration, it is assumed that the process of schema re-creation begins with an external stimulus.

Note also that some units have a higher likelihood of being activated because the feature represented by a unit occurs more frequently in the external environment. Similarly, some units are activated less frequently because the feature rarely occurs. Connectionist models of schemata normally introduce a bias into the model to account for this phenomenon. If a unit is usually activated, a positive bias is introduced; if a unit is not normally activated, a negative bias is introduced.

AN APPLICATION: A MODEL OF SCHEMATA FOR AUTOMOBILES

One application of schemata theory involves the features or attributes normally associated with a product. Meyers-Levy and Tybout (1989) identify a number of attributes of various beverages; Sujan (1985) considered the attributes associated with a 35mm camera. An automobile might also be represented by a group of attributes. Some attributes are concrete (e.g., power windows), while others are more abstract (e.g., high status). Table 1 presents a listing of 20 attributes that might be associated with the automobile product category. These attributes were chosen for their diversity in concreteness/abstractness and for their ability to represent several types of automobiles. Note that some properties of automobiles (e.g., price) are shown as two attributes (e.g., high price, low price), since each unit in this simple constraint satisfaction network is binary (either on or off) and cannot represent multiple levels of continuous variables.

The Model

A connectionist approach to schema retrieval can be modeled mathematically. The model must update the activations on each of the 20 units in the system based on the inputs from other units, any external input, and the constant bias. Rumelhart et al (1986) provide the following model:

EQUATION

where aj(t) represents the activation of unit j at time t, an internal input is represented by setting aj(t)=1, netj(t) represents the net input to unit j at time t, and

EQUATION

where ej(t) represents the bias, wji is the weight (constraint) on the connection from unit i to unit j, and ai(t) represents the activation of unit i at time t.

The system will reach a stable state with all units either turned on (a value of 1) or turned off (a value of 0). If the net input to a unit at time t exceeds 0, then the activation on that unit will be driven toward 1; otherwise, its activation will be driven toward 0. Note that this is an iterative model, and that only one unit may be updated at a time (although a truly parallel processing system would allow simultaneous updates). The model randomly chooses one unit at a time to update activation; this unit may not be selected again for updating until the remaining 19 items have been updated. Each cycle of updating all 20 units is considered one complete update.

Setting Connection Weights

The weights in the model represent the connection strengths between units. The weights will be positive when two units are mutually excitatory and negative when two units tend to inhibit each other. To obtain weights meeting these criteria, data were collected from 44 undergraduate students. Each subject completed a written questionnaire; all questionnaires were identical except for the order of the stimulus materials. Initially, subjects were presented with a list of the 20 attributes of automobiles shown in Table 1 and were instructed to indicate all attributes that applied to a general type of automobile. For example, approximately one-third of the subjects were told to consider a LUXURY SPORTS CAR and "to think of the typical automobile of this general type" and to complete the page by marking each attribute that they "believe might in general apply to the automobile." Approximately one-third of the subjects began by considering the typical ECONOMY CAR, while the remainder initially considered the typical FAMILY CAR. When subjects had considered all 20 attributes for the first type of automobile, they repeated the process for the two remaining types of automobiles. Thus, each subject considered all three types of automobiles, for 132 total observations. The order of the presentation of the three types of automobiles was balanced to eliminate order effects.

TABLE 1

WEIGHTS (CORRELATIONS)

Correlations were calculated across all observations for all pairs of attributes to obtain weights for use in the constraint satisfaction model. An attribute was coded "1" if a subject selected it; otherwise, the attribute was coded "0". Although the range of the correlations obtained using two dichotomous variables may be restricted, depending on the shapes of the distributions of the two variables (Nunnally 1978), the correlations obtained are useful measures of association. These correlations are shown in Table 1. For example, a leather interior is positively correlated with a high price (r=0.77); in the model, the unit representing leather interior is therefore connected to the unit representing high price by a connection weight of 0.77. Note that weights for four of the combinations (high price-low price; high status-low status; foreign-made-American-made; and, often needs repair-rarely needs repair) were set to -1.00.

Calculating Bias

In addition to the information used to estimate weights, subjects also provided information used to calculate the biases needed in the model. Subjects were asked to estimate for each attribute the percentage of cars that exhibit that attribute. These percentages are presented in Table 2, along with the biases for each attribute obtained by applying the following formula from Rumelhart et al (1986):

EQUATION

Model Summary

Twenty attributes of automobiles were specified, each represented by a single unit; each unit represents the hypothesis that the feature is present and has value "1" when the model specifies that a feature should be present and value "0" when a feature should not be applicable. All units are connected to all other units and receive positive or negative impulses based on the connection weight and the current activation level of the impulse-sending unit. When the stable state is achieved, all units have activations of 0 or 1; the units that are activated represent the relevant schema.

A SCHEMA FOR HIGH STATUS AUTOMOBILES

Suppose that this system received strong external evidence that the hypothesis related to "high status" was correct, and that the activation on this unit was accordingly set to 1 (and not allowed to change). How might the system respond? The unit that is activated begins to send positive impulses to those units with positive connection weights and inhibitory impulses to those with negative connection strengths. The system will ultimately settle into the best possible set of unit activations that it can achieve, given the constraints between units. This situation was modeled and is shown in Table 3. The entries in the table represent the activation levels of each unit at the conclusion of a complete update (e.g., the activation of the first unit, foreign-made, is equal to 0.25 once the model has completed the first update). Although the model required 105 updates to reach the stable state, only the first five and the last updates are shown. The model required updates 6-105 to bring the activation of unit 16 (advertised on national TV) to 0. This complete process was repeated 20 times to obtain a measure of the stability of the schema; this pattern was obtained 19 times. The schema for a high status automobile suggests an expensive, foreign-made automobile that might be special ordered from the factory, one that rarely needs repair, and has the following features: power windows, leather interior, and power steering. In addition, the automobile serves the basic purpose of providing transportation. This schema seems to reflect several of the attributes that might normally be associated with high status automobiles.

It is interesting to note the progression of the activity levels of individual units as the model proceeds through the update process. Some of the units (e.g., power steering and provides transportation) receive full activation on the first update while others (e.g., foreign-made and rarely needs repair) start slowly, building activation with each update. Many units which are initially off continue to remain off throughout the process. Other units receive some positive activation along the way but ultimately decrease to zero activation. The intermediate activations reflect the interactions among the units as they approach the stable state.

A SCHEMA FOR LOW STATUS AUTOMOBILES

It is interesting to next determine the resulting schema when unit 11 (low status) receives strong external stimulation. This external input was implemented and the model was allowed to run until a stable state was attained. Table 4 presents the ending activations for each of the six updates required to reach the stable state. The resulting pattern of activations was obtained each time that the simulation was run (20 times). A low status car might be described as an inexpensive American-made car with vinyl seats, a roomy interior, a small engine, and an AM radio; the car often needs repair (but parts are readily available). The car provides transportation and is advertised on national TV. It is interesting to note that this schema appears to be the "typical" automobile schema, since it represents the final stable state of the system when no strong external inputs are assumed (i.e., the system is allowed to run using the biases as the only inputs to the system). The system was allowed to run 20 times with all initial activations set to zero; in 19 cases, this schema resulted.

TABLE 2

BIAS ESTIMATION

TABLE 3

CONSTRAINT SATISFACTION SCHEMA, ITEM 4 ACTIVATED

TABLE 4

CONSTRAINT SATISFACTION SCHEMA, ITEM 11 ACTIVATED

OTHER EXAMPLES: COMBINATIONS OF EXTERNAL INPUTS

In many situations, it is likely that more than a single unit may receive strong external stimulation. Consider the situation in which the units representing "high status" and "American-made" both receive strong external evidence that the attributes exist. To demonstrate this situation, the activations on these units were set and held at 1; the model was then allowed to update in the normal way. The resulting schema was less stable than the earlier examples; two general schemata resulted. The first, which was produced on 11 of the 20 simulations, is shown in Table 5. The model took 16 iterations to reach the stable state. The resulting schema was of a high priced, high status, American-made automobile with features in accordance with its status and price. On 9 of the 20 simulations, the model produced a less grand schema: the automobile provides transportation, has a roomy interior, vinyl seats, power steering, an airbag has readily available parts, and is advertised on television.

DISCUSSION

Following Rumelhart et al (1987), several important properties of schemata are easily addressed within the connectionist approach. Several of the properties relevant to consumer research are broadly considered in this section.

(1) Patterns tend to complete themselves. In order to retrieve a schema, it clearly is not necessary to receive external input for all units. As demonstrated, the interconnected system of constraints serves to activate those units, or attributes, that fit the schema. For example, when a "high status" automobile was indicated, the model settled into a solution with attributes that look very much like those appropriate for a high status automobile.

As another example, Sujan et al (1986) found that the stimulus of a salesperson was sufficient for people to bring up the clothing store salesperson schema. It was not necessary to get external input on such units as "...will tell me I look good in anything I've chosen to try on" or "...is probably selling on commission". Rather, the system settled into a solution in which these or other appropriate units were activated.

Much of the rationale behind brand extensions relies upon this principle. By attaching a well-known brand name to a new product, the marketer assumes that many of the units normally activated in the consumer's schema for the brand name will also be learned by the system in association with the new product. In this way, consumers may attach many associations to the new product as a result of the brand schema, without an extensive (and expensive) effort by the marketer to build the desired associations.

TABLE 5

CONSTRAINT SATISFACTION SCHEMA, ITEMS 4 AND 10 ACTIVATED

They result simply through the process of re-creating the schema for the brand.

(2) Knowledge at all levels can be represented in schemata. In the automobile example, some of the attributes were very concrete, while others were very abstract, yet the model adequately represents all levels of abstraction. Consumers can thus be expected to incorporate higher level abstraction into their schema for a certain product. The image of a brand may often be represented in the brand schema along with more tangible elements of the physical product. For example, abstract advertising for Obsession perfume has been used to convey an image of the product.

(3) Schemata are active processes, not static patterns stored in memory. In this perspective, schemata are not stored at all; rather, the connection strengths between units representing hypotheses about features or attributes are retained. Over time, through the processing of schema-related information, the connection strengths may be altered. Advertising and promotion may serve to adjust some of the connection strengths between units. Advertising for a new product may be thought of as an attempt to increase the connection weights between a unit representing the product and units that tend to form a schema of products to consider for purchase (i.e., the consideration set). Similarly, connectionism provides an alternate explanation of the process through which a product is assimilated into an existing product class than does the schema-plus-tag explanation offered by Sujan and Bettman (1989).

One intriguing potential use for a connectionist model of schemata is the identification of specific schema elements or units to serve as the foci for advertising efforts. Suppose that a company offers a brand that is currently viewed as having low status and that the company desires to change this perception. With knowledge of the various weights between units in the system, it may be possible to identify a specific unit or attribute that, when activated, consistently activates the high status unit in the stable state. An advertiser could then focus communication efforts on this attribute. If the attribute does not currently exist for the brand, it may be added. It is possible that the addition of an inexpensive attribute might lead to dramatic changes in consumer perceptions of the product.

(4) Some schemata are more rigid than others. The units of some schemata seem to be much more tightly connected than those of other schemata. For example, the general schema for high status automobiles shown in Table 3 seems to be much more stable than does the schema for high status, American-made automobiles shown in Table 5. The high status schema is clearly defined, having resulted 19 out of 20 times it was modeled. However, individuals appear to hold a less clear schema for high status American cars since this particular schema is produced only about half the time.

The schemata formed for certain products (where the schemata might describe the attributes of the products at a basic level, or the positioning of the products at a higher level of abstraction) seem to be very tight and rigid, while schemata for other products are much less so. For instance, consider the schemata formed for IBM personnel computers in comparison with schemata for Zenith computers. It is likely that the position of the IBM product is very clear and tightly formed, while that of the Zenith product is less rigid. One implication is that tight schemata might be more difficult to combat (from a competitive strategy perspective) or change in consumers' minds. This implication can be investigated from a connectionist perspective.

With this advantage for the company or brand with the strongest position, however, comes the possible disadvantage of inflexibility. For example, Coca-Cola's ill-fated attempt to change the product formula may have failed to a large extent because individuals held a clear conceptualization of what the product was "supposed" to be. New Coke may have simply been too great a departure from the tightly-held schema for the product.

CONCLUSION

A constraint satisfaction model of schemata based on connectionism has much to offer consumer research as a theoretical underpinning for research on schemata. In this paper, a brief overview of such a model has been presented and demonstrated. The results indicate that the model seems to capture the nature of consumer schemata for automobiles.

REFERENCES

Ashcraft, Mark H. (1989), Human Memory and Cognition. Glenview, IL: Scott, Foresman and Company.

Cohen, Joel B. and Kunal Basu (1987), "Alternative Models of Categorization: Toward a Contingent Processing Framework," Journal of Consumer Research, 13(March):455- 472.

Hintzman, Douglas L. (1990), "Human Learning and Memory: Connections and Dissociations," Annual Review of Psychology, 49:109-139.

John, Deborah Roedder and Mita Sujan (1990), "Age Differences in Product Categorization," Journal of Consumer Research, 16(March):452-460.

Loken, Barbara and James Ward (1990), "Alternative Approaches to Understanding the Determinants of Typicality," Journal of Consumer Research, 17(September):111-126.

McClelland, James L., David E. Rumelhart, and Geoffrey E. Hinton (1986), "The Appeal of Parallel Distributed Processing," in J.L. McClelland, D.E. Rumelhart and the PDP Research Group (Eds.), Parallel Distributed Processing, (Volume 1), pg. 3-44. Cambridge, MA: MIT.

Meyers-Levy, Joan, and Alice M. Tybout (1989), "Schema Congruity as a Basis for Product Evaluation," Journal of Consumer Research, 16(June):39-54.

Nunnally, Jum C. (1978), Psychometric Theory, (Second Edition). New York: McGraw-Hill.

Rumelhart, David E., Paul Smolensky, James L. McClelland, and Geoffrey E. Hinton (1986), "Schemata and Sequential Thought Processes in PDP Models," in J.L. McClellend, D.E. Rumelhart and the PDP Research Group (Eds.). Parallel Distributed Processing (Volume 2), pg. 7-51. Cambridge, MA: MIT.

Sujan, Mita (1985), "Consumer Knowledge: Effects on Evaluation Strategies Mediating Consumer Judgments," Journal of Consumer Research, 12(June):16-31.

Sujan, Mita and James R. Bettman (1989), "The Effects of Brand Positioning Strategies on Consumers' Brand and Category Perceptions: Some Insights from Schema Research," Journal of Marketing Research, 26(Nov):454-467.

Sujan, Mita, James R. Bettman and Harish Sujan (1986), "Effects of Consumer Expectations on Information Processing in Selling Encounters," Journal of Marketing Research, 23(Nov):346-353.

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