Network Analysis of Autobiographical Memory: Investigating Small-Worldness in Memory and Feature Networks
Neele C. Kinkel University of Amsterdam
Abstract
Memory theorists posit that autobiographical memories are stored in an autobiographical memory network. However, this network has never been assessed. This study primarily focused on developing a method for investigating the autobiographical memory network. The exploratory aim of this study was to detect if there is small-worldness in the memory and feature networks. The participants were 12 students from the University of Amsterdam. Over the period of 5 days, they recalled 100 autobiographical memories and 20 future
imaginations. We pulled out the salient features of the memories and established connections between memories with shared features and features that occur in the same memories. We used the clustering coefficient and average shortest path length of the observed and random network to calculate the small-world index for each participant’s memory and feature network. The memory networks did not exhibit small-worldness, but the feature networks did. The current study helps understanding the accomplishment of autobiographical memory storage and retrieval.
Network Analysis of Autobiographical Memory: Investigating Small-Worldness in Memory and Feature Networks
Autobiographical memory plays a significant role in our everyday lives. It can be defined as the ability to consciously remembering personal events (Berntsen & Rubin, 2012; Williams, Conway & Cohen, 2008). An autobiographical memory can be any memory of the past: it can be of a major event or a minor event, a long time ago or quite recently.
Autobiographical memory guides how we perceive ourselves (Conway, 2005). It helps us understanding our position in a complex social environment, where it enhances orientation and participation (Berntsen & Rubin, 2012). Therefore, autobiographical memory contributes to our lives every single day.
Within autobiographical memory a distinction can be drawn between semantic and episodic memory (Tulving, 1972; Greenberg, 2010). Semantic memories are facts about the world, general knowledge, abstractions and generalizations, often learned from others (Williams, Conway & Cohen, 2008), e.g. “the main ingredient of hummus is chickpeas.” They can be about the self, for example “I like hummus,” but do not concern specific events from one’s life. On the other hand, episodic memories are memories about specific actions, action outcomes, people, objects or events (Conway, 2005; Williams, Conway & Cohen, 2008), e.g. “I was in France, when I made my own hummus for the first time.” They can be considered as mental time-travelling, i.e. remembering an event experienced by oneself (Conway, 2005). In this study, we will focus on episodic memories.
Each of us carries an incredibly large amount of episodic memories and it is astounding that we are able to effectively store and retrieve them (Wagenaar, 1986). Even when consulting only daily routines like drinking coffee, eating lunch or sitting in a lecture, a rich set of specific memories pops up. Now consider, given the amount of single days – or even hours – we have lived that were full of unique experiences, the diversity of places we
have visited and people we have been in contact with, how many autobiographical memories our brain comprises. Learning about the organization of autobiographical memory can help us understand how we are able to accomplish the feat of memory recollection. Theories about the organization of autobiographical memories have been developed, trying to fill in the larger picture of memory recollection. In the following paragraphs, I will explain, analyze and evaluate two leading theories about autobiographical memory organization.
According to a theory by Conway and Pleydell-Pearce (2000) autobiographical memories can be seen as transitory mental constructions within a so-called self-memory system. This self-memory system has two principle components: the autobiographical knowledge base and the working self. The autobiographical knowledge base contains three levels of specificity: lifetime period knowledge, general event knowledge and event-specific knowledge. Lifetime period knowledge is knowledge about longer periods of one’s past life, for example “the time I did my BA in Amsterdam” or “the time I worked in the hummus restaurant,” as well as the duration of those periods. It can be ordered in different themes, for example “career theme,” “relationships theme,” etc. General event knowledge includes more specific details than lifetime periods, like, within the “career theme,” the memory of a team meeting or a job interview. It summarizes repeated or single events of different lifetime periods, so it gives access to less contextual and more distinct knowledge. General events are connected with each other, but also with the lifetime periods. At this level, the memories might form small thematically related groups with a distinct local organization. Finally, event-specific knowledge includes detailed, often sensory-perceptual, information – images, feelings, smells, etc. – about a specific event experienced in an individual’s life, like “the first time I went to Israel and ate real hummus.” In other words, event-specific knowledge
includes the most important features of a specific memory that are not covered by general event knowledge, like where exactly an experience took place, who was there, etc. These
three levels of autobiographical knowledge are organized hierarchically: event-specific knowledge is embedded in general event knowledge which can in turn be placed more broadly in lifetime periods.
According to this theory, a memory is a piecewise construction that draws on the information from this knowledge base at the different levels of specificity: it incorporates broader information from the lifetime periods, more explicit information from the general event knowledge and unique information from the event-specific knowledge. The recall, as well, is never solely of one type of knowledge: we recall components of memories along the different levels of the autobiographical knowledge base. Access to the memories is guided by the so-called “working self,” which is a set of hierarchically structured goals. It constraints access to memories that are not coherent with the personal goals and facilitates access to memories that are indeed coherent with personal goals. According to this theory,
autobiographical memory has a multiple nested structure: a memory is stored in different memory traces, so that different levels of the memory can be remembered independently (Neisser, 1986).
The theory of Conway and Pleydell-Pearce (2000) is generally supported by the literature. Eade et al. (2006) conducted five studies to test the hierarchical model of autobiographical memory retrieval among healthy participants. They found that memory recollection began with broad general events before digging into specific recollections. When stopped before terminating autobiographical memory retrieval, participants often remained on a general level, instead of recollecting specific details. The results supported the basic idea of hierarchical memory retrieval in the autobiographical knowledge base (Conway &
Pleydell-Pearce, 2000), that suggests that memory recollection begins broadly and ends specifically. Haque et al. (2014) also found support for the hierarchically structured search strategy among a non-clinical sample and depressed participants. The results confirmed that
both non-clinical and depressed participants start memory retrieval at lifetime periods and move further to general event knowledge. Depressed participants terminated their search at the general event level, while healthy participants proceeded to event-specific knowledge. Both studies support basic ideas of the hierarchical model. However, the given cues were held at a general categorical level (i.e. common locations, objects or emotions) (Haque et al., 2014), which could alternatively explain why participants started memory recall at the most general level (lifetime periods) before they moved to the more specific level(s). Berntsen and Rubin (2012) criticized the narrow focus on the spatial dimension in nested autobiographical memory models. Because the self-memory system theory only allows abstractions along the temporal dimension, it is characterized by inaccurate simplicity, according to Berntsen and Rubin (2012).
Therefore, Berntsen and Rubin (2012) adopted a different approach about the
structure of autobiographical memory. Instead of thinking of it as a stable and hierarchically structured system, they propose a flexible structure which allows autobiographical knowledge abstraction among different dimensions. The researchers identify the temporal, spatial and social dimension which can be considered salient features of a memory. Which of those dimensions dominates the structuring of autobiographical memory depends on how central their role is in the organization of a person’s past. For some memories, like the 18th birthday, the temporal dimension might be more outstanding, while for others, like a holiday, the spatial dimension might stand in the foreground. Within the dimensions, memories are also organized by their level of abstractness. Abstraction among the temporal dimension means that we can memorize events of a wider timespan, like “the time when I biked to university every day,” or of a specific day, like “the day when I fell from my bike.” The same
too, autobiographical memory yields a multiple nested structure, but that structure is flexible and varies as a function of the individual and their current context.
Different evidence exists in the literature for autobiographical memory organization structured by each of the three dimensions identified by Berntsen and Rubin (2012). For example, Brunec et al. (2015) found that when participants were to recall memories that happened shortly before or after a salient event, they were structured in a chronological order, which provides support for the existence of the temporal dimension. However, this structure along the temporal dimension was only proven for memories from a restricted time scale of a few days, but not for weeks or months. Sheldon and Chu (2017) found that the retrieval of autobiographical memories, cued by the spatial dimension, for example “restaurant,” costs participants less time compared to retrieval cued by a thematic dimension, for example “family get-togethers,” which provides support for the existence and occasional dominance of the spatial dimension. At the same time, their results undermine the notion that “lifetime periods” are the sole organizing dimension as the self-memory system theory proposes (Conway & Pleydell-Pearce, 2000). Berntsen and Rubin (2004) demonstrated that culturally shared life scripts of social roles and social settings play a role in recall of emotionally loaded autobiographical memories. Therefore, the social dimension might play a role in
autobiographical memory organization. All in all, the literature does support the existence of each of the dimensions proposed by Berntsen and Rubin (2012). However, this theory does not clarify in what way the organization of autobiographical memory is supposed to be “flexible.” That is, it does not specify how the different dimensions are related and what exactly determines the salience of features.
In order to test the theories by Conway and Pleydell-Pearce (2000) and Berntsen and Rubin (2012) it is necessary to map the structure of the autobiographical memory network, which has never been done. Fortunately, there is a whole interdisciplinary field devoted to
understanding the global organization of systems: network science (Barabási, 2012). It is already widely used in modelling of physical, biological and social phenomena (Barabási, 2011). In the – by every minute growing – complexity of autobiographical memories, network analysis can not only help us to identify a structure, but also to understand their storage and retrieval. A network consists a set of elements (nodes, N) and the relationships between them (edges, K) that connect them. Every node has a degree, which is the number of other nodes it is connected to. For example, consider the train system of the Netherlands. The nodes are the train stations, the edges are the rails and the degree of a station is the number of train stations that could be reached from that station without passing another station. A
station with a very high degree would be considered a “hub” in the network (Barabási, 2012). Network science has shown that many networks from different areas have shared structural properties. For example, many complex networks display small-world properties (Borsboom et al., 2011; Watts & Strogatz, 1998; Watts & Bullmore, 2006). The worlds social network illustrates the basic idea of a small-world network: we are connected to any other person in the world through six or less “handshakes,” i.e. people we know. The nodes of this network are the people in the world and the edges are the handshakes. This so-called “six degrees of separation”-theory (Barabási, 2012) demonstrates that, despite long distances and community extremities, we are able to detect a certain overarching structure of the world’s population.
Of particular relevance to the current study, small-worldness has been found in semantic memory networks. Morais, Olsson and Schooler (2013) looked at associative networks of semantic memories. During six weeks of daily 1-hour sessions, six participants were asked to give associative responses to cue words, for example “angry” or “adult.” The responses were used as cues in the next session. This procedure was repeated over five to seven sampling iterations, so the associative semantic networks grew like a rolling snowball.
Words were considered linked if they were recalled as response to the cue word. The two most important characteristics of small-world networks – high local clustering and a short average path length – were detected in the built networks. Clustering can be described as a group of nodes being closely together. Local clustering looks at the level of clustering behavior on certain locations in the network. A short average path length means that all elements in the network can be reached by any other element in a relatively small number of steps (Bassett & Bullmore, 2006). Hills, Todd and Jones (2012) looked at how we move within semantic memory networks, comparing static and fluid patch models. According to fluid patch models, searching for a new word depends on the last recalled word. According to static patch models we switch from subcategory to subcategory only after depletion of one subcategory. The results supported the fluid patch models, implying that searching in
semantic memory is not random: semantic memory networks can be described as patchy and we can easily reach other nodes within the network. Knowing that semantic memory has a small-world network structure (Morais, Olsson & Schooler, 2013) and that we effectively move from cluster to cluster within these networks (Hills, Todd & Jones, 2012) helps us understanding the basic principles of semantic memory.
If small-worldness is also observed in the autobiographical memory network, it could contribute a large part of understanding how we are able to effectively store and retrieve autobiographical memories. However, episodic memories have never been mapped out as a network. The primary aim of this study is to develop a method of investigating the
autobiographical memory network. The current study functions as a pilot study to begin this process. We will then use this network to conduct a preliminary examination of whether these networks exhibit small-worldness.
Methods Participants
70 participants signed up for the project and completed the informed consent. 42 participants started the first questionnaire. 11 of them did not finish it and another 5 dropped out after completing. 12 participants dropped out before filling in the fifth questionnaire. 14 participants completed all 5 days of the autobiographical memory questionnaire. 2
participants were omitted, because their answers were in Dutch. Our final sample contained 12 participants. All participants were 18 years or older. They were first-year psychology students at the University of Amsterdam, who read about the research in the research
participant recruitment section of their university. They received 3.75 points for participation. There were no exclusion criteria for participation. Participation was voluntary and
anonymous. Procedure
When registered for participation on lab.uva.nl, participants received an e-mail by the research team, in which all information was outlined. The email included information about the time span of the research (5 hours), the online environment of the questionnaire, the reward of participant points and the general procedure of getting daily e-mail reminders until all answers were recorded. The link to the questionnaire in Qualtrics was attached in the first e-mail, so participants could start right away. During the next 4 days, participants received daily e-mails that included the link to the questionnaire. In case participants missed to fill in the questionnaire for one or two days, additional reminder e-mails were sent to them until the data collection was completed.
Materials
The questionnaire was administered online. All answers were recorded in Qualtrics (Qualtrics, 2018). Participants could access the questionnaire via a link that was sent to them by e-mail. When following the link, participants first had to complete the informed consent. In the questionnaire, participants were asked to provide 20 autobiographical memories, 3 of
which were self-defining memories, as well as 4 future imaginations, needed for a broader project. Self-defining memories are important, personal memories that help individuals most clearly to define how they see themselves (Singer & Moffitt, 1992; Thorne & McLean, 2001). Future imaginations are imaginations of events that may occur in the personal future (Tulving & Murray, 1985). Participants were requested to share their autobiographical and self-defining memories, as well as the future imaginations, in an open response format with a maximum of four sentences. Additionally, they were asked to share the approximate date (year, month and day), place (country, city and/or specific location), people who were there (first names and initial(s) of last names), emotion (multiple choice of basic emotions) and valence (100-point likert scale from extremely positive to extremely negative). There were additional subquestions for the self-defining memories and future imaginations, that were not relevant to this study.
Data Analysis Plan Data Preparation
We downloaded the raw data from Qualtrics as csv-files and read them into RStudio (RStudio, 2018). There, we conducted three main steps to pull out the salient features of the memories. Firstly, from the recorded data, we pulled out the critical components that contain potential features of the memory: memory descriptions, the most specific location variable (e.g. “in a church”) and the persons who were in the memory. Secondly, using the R-package “tokenizers” (Mullen et al., 2018), we broke this information up into individual components, so they became separate words. For example, “I went home” became “I,” “went” and
“home.” With the same package, we modified the components, removing capitalization and punctuation: “I,” “went,” “home” became “i,” “went,” “home.” In addition, for people and places, we eliminated spaces in order to create complete unique identifiers of these features, e.g. “Tom T” became “tomt” and “my house” became “myhouse.” Thirdly, using the “tm”
package (Feinerer & Hornik, 2017), we removed components that we did not consider features (i.e., words that did not provide information about the memory itself), such as stop words (Benoit, Muhr & Watanabe, 2017), prepositions (Essberger, 2012), amplification words, deamplification words, negation words, adverbs, function words (Rinker, 2018), pronouns (“List of Pronouns in English”, 2018), adjectives (“Lists of Adjectives”, 2018) and auxiliary verbs (“Auxiliary Verbs”, 2018). Afterwards, we combined all extracted features in one vector. The aim was to keep only verbs and nouns as salient features of the memory descriptions. Table 1 walks through the four steps of the data preparation.
Table 1
Three Steps of Data Preparation
Steps Example
1: Pull out features I was on holiday in France. After a long car trip, when we finally arrived to the house, we cooked some hummus and drank Gin Tonics.
2: Break up and modify
[1] "i" "was" "on" "holiday" "in" "france" "after" [8] "a" "long" "car" "trip" "when" "we" "finally" [15] "arrived" "to" "the" "house" "we" "made" "cooked" [22] "some" "hummus" "and" "drank" "gin" "tonics"
3: Remove words [1] "holiday" "france" "car" "trip" "finally" "arrived" "house" [8] "made" "cooked" "hummus" "drank" "gin" "tonics"
By repeating these data preparation steps for each memory, we produced a two column matrix for every participant with column one including the 100 memories and column two including the extracted features. This matrix could be considered an edge list.
Table 2 shows an example of an edge list. From here, we could go on to the actual formation of the networks.
Table 2
Edge List Example Memory Features M1 myhouse M1 myfriend M1 played M1 game M1 noamr M1 won M2 i M2 sad M2 hamster M2 died M2 ate M2 ice-cream M3 i M3 spain M3 vacation M3 enjoyed M3 sun M3 tilk M4 myhouse M4 tired
M4 longday
M4 bed
M4 noamr
Network Formation
The network formation was divided in two broader steps: (a) building a two-mode network and (b) using that network to construct two one-mode networks for every
participant. Two-mode networks are also known as bipartite networks and they contain two sets of nodes, which were in our case the memories and the features of the memories. In the two-mode networks, ties are only established between nodes belonging to different sets. In our case ties were established between memories and memory features, but not between memories. From the two-mode networks two one-mode networks were derived: one of them had the memories as nodes, while the other one had the memory features as nodes. Figure 1 shows an example of a bipartite network and its projections. In the projection process, one set of nodes (e.g. memories) is chosen and nodes are connected, if they are both connected to at least one of the same nodes of the other set in the two-mode network (e.g. if two memories are both connected to a given memory feature). The number of shared nodes of the other set defines the weight of the connection, i.e. if memory 1 and memory 2 are both connected to feature 1 and 2, the weight of the connection between memory 1 and memory 2 equals 2.
Figure 1. Two-mode network and one-mode networks. A bipartite network with memories and features (middle figure), a one-mode memory network (left figure) and a one-mode feature network (right figure).
Network Analysis
We assessed the density of the networks, that ranges from 0 (very sparse) to 1 (dense) and captures how many of the possible connections actually existed in the observed
networks. Small-worldness was assessed through the small-world index (Humphries, 2008), calculated through the clustering coefficient (transitivity) and shortest average path length for every network (Borsboom et al., 2011). The clustering coefficient ranges from 0 to 1 and measures to what extent the neighbors of a node are connected to one another themselves. For example, when A is connected to B and to C, and B and C are connected as well, A has a clustering coefficient of 1. Shortest average path length is an average of all the network’s shortest paths (i.e. the minimum number of edges that must be traversed to move from node A to node B). For example, if the shortest average path length in a network is 2, on average two edges have to be passed over to get from A to B.
In order to calculate the small-world index, we also looked at random networks. In a random network, the links between nodes are placed randomly: each pair of N nodes is connected with probability p. For every participant, we calculated the clustering coefficient of the observed network ( ) and a random network ( ), as well as average path lengths in the observed network ( ) and a random network ( ). The small-world index thus captures what was observed relative to what would be seen in a random network. We calculated it through:
We handled the conservative worldness criteria threshold of 3, i.e. a small-world index above 3 indicates the presence of small-small-worldness (Humphries, 2008). We looked at the small-world indices of the constructed memory networks and feature networks.
Results
All memory networks had 100 nodes, which were the memories every participant reported, and between 1092 and 2880 edges between memories (M = 2002.92, SD = 525.30). Across networks, the mean average degree was 40.06 (SD = 10.51), which means that a node was on average connected to 40 other nodes. The memory networks exhibited low to medium density (M = .40, SD = .11): of all possible connections between the reported memories, approximately half of them were actually present in the observed networks. Figure 2 shows the resulting autobiographical memory network for participant 1.
Figure 2. Memory network of participant 1. The bigger and darker a node, the higher its clustering coefficient. Layout by Fruchterman & Reingold (1991).
Table 3 shows the results of the analysis of the memory network for our primary aim. On average, the observed networks had a clustering coefficient (transitivity) of .60 (SD = .07), which was a bit higher than the clustering coefficient in a random network (M = .40, SD = .11). The nodes in the observed network were more linked to each other than they would be in a random network. The shortest average pathlength in the observed networks was on average 1.62 (SD = 0.14), which was approximately the same as in a random network (M = 1.61, SD = 0.09). It means that on average 1.62 edges needed to be passed over to reach another memory and this was approximately the same in the observed networks and in random networks. Across all participants, the small-world index was consistently smaller than 3 (M = 1.56, SD = 0.28).
Table 3
Small-world properties of the memory networks Participant Transitivity in observed network Transitivity in random network Shortest APL in observed network Shortest APL in random network Small-world index 1 0.54 0.22 1.95 1.81 2.31 2 0.52 0.29 1.73 1.70 1.76 3 0.63 0.43 1.57 1.57 1.44 4 0.61 0.40 1.60 1.59 1.51 5 0.60 0.34 1.70 1.65 1.71 6 0.63 0.45 1.55 1.56 1.41 7 0.52 0.31 1.70 1.67 1.65 8 0.72 0.58 1.42 1.49 1.32 9 0.64 0.41 1.59 1.58 1.53 10 0.54 0.40 1.59 1.59 1.34 11 0.72 0.57 1.42 1.50 1.33 12 0.59 0.40 1.60 1.59 1.46 Mean 0.6 0.4 1.62 1.61 1.56
The feature networks had between 401 and 990 nodes (M = 716.33, SD = 161.85) and between 2640 and 14576 edges (M = 9146.92, SD = 3480.28). The nodes had an average degree of 24.63 (SD = 5.10), which means a node was on average connected to 25 other nodes. All feature networks exhibited very low density (M = .03, SD = .00), which means
they were sparse networks. Only a very small percentage of all possible connections between the reported memory features was actually connected in the constructed networks. Figure 3 shows the resulting feature network for participant 1.
Figure 3. Feature network of participant 1. The bigger and darker a node, the higher its clustering coefficient. Layout by Fruchterman & Reingold (1991).
Table 4 shows the results of the analysis of the feature network. The clustering coefficient in the observed feature networks was on average .33 (SD = .02), while in the random networks it was only .03 (SD = .00). Thus, there was more clustering in the observed feature networks than in the random networks. The shortest average path length in the
observed feature networks was on average 2.40 (SD = 0.12), while in the random networks it was 2.38 (SD = 0.09). On average, 2.4 edges needed to be passed over to get from one feature to any other feature observed feature networks. This was approximately the same in the
random networks. Across all participants, the small-world index was consistently larger than 3 (M = 9.59, SD = 1.42). Thus, all feature networks exhibited small-worldness.
Table 4
Small-world properties of the memory feature networks Participant Transitivity in observed network Transitivity in random network Shortest APL in observed network Shortest APL in random network Small-world index 1 0.33 0.03 2.70 2.60 9.97 2 0.35 0.03 2.47 2.42 10.84 3 0.34 0.04 2.38 2.36 9.20 4 0.31 0.04 2.39 2.40 8.44 5 0.32 0.04 2.44 2.39 8.25 6 0.33 0.03 2.34 2.32 9.71 7 0.36 0.03 2.50 2.42 11.90 8 0.31 0.04 2.22 2.27 7.88 9 0.35 0.03 2.42 2.43 11.40 10 0.33 0.03 2.40 2.37 10.86 11 0.28 0.04 2.21 2.32 7.70 12 0.34 0.04 2.33 2.31 8.88 Mean 0.33 0.03 2.40 2.38 9.59 Discussion
The primary aim of this study was to develop a method of investigating the autobiographical memory network. The exploratory aim of this study was to detect small-worldness in the memory and feature networks. We found small-worldness in the feature networks. In the memory networks, however, we did not find small-worldness.
The hierarchical organization discussed in both theories suggested that there should be significant clustering at the feature level. According to Conway & Pleydell-Pearce (2000), event-specific knowledge (features) should be organized in clusters, as they are embedded in the themes of general event knowledge which are in turn organized along the lifetime periods. Cues held at the general event knowledge level then activate regions (clusters) of event-specific knowledge. In the flexible hierarchical organization proposed by Berntsen and Rubin (2012), we should see clusters of the temporal, spatial and social dimension at the most concrete (feature) level, as well as at the more abstract levels. We found clear evidence of clustering in the feature network: the features form small interconnected groups within the whole network. This seems consistent with the discussed theories. Thus, memory features seem to be regionally specialized in certain periods, like a computer network where different machines carry out certain assigned tasks (Telesford et al., 2011).
Furthermore, the small-world index of the feature networks shows us that there is both clustering and short average path length: the features are not only organized in small distinct groups, but at the same time any given feature is close to any other feature. The short average path length allows us to easily get from a feature of a childhood memory (e.g. the smell of our old garden) to a feature of a memory from yesterday (e.g. the taste of the hummus I ate yesterday), even though those features might be located in different clusters. Like a social network, where there are interconnected groups of friends, but each person is really only six handshakes away from everybody else, memory features exhibit a structure that may allow for effective storage and retrieval of memory features (Hills, Todd & Jones, 2012).
Both theories also suggested that there should be clustering at the memory level. Conway and Pleydell-Pearce (2000) proposed some clustering at the general event knowledge (memory) level, although less than at the event-specific knowledge level, while Berntsen and Rubin (2012) proposed clusters of the temporal, spatial and social dimension at all levels of abstraction, including the memory level. We did see evidence for clustering in the memory network. As substantial clustering was also observed in the random network, much, though not all, of the clustering in the memory network could be accounted for by the density of the network. The shortest average path length in the memory network was low, which allows to easily get from the memory of my sixth birthday to the memory of eating hummus yesterday.
The clustering that we found in both the memory and the feature network provides substantial support for the theories of Conway and Pleydell-Pearce (2000) and Berntsen and Rubin (2012). As posited by Conway and Pleydell-Pearce (2000), the abstraction along the temporal dimension suggests that there should be more clustering on the lowest level of organization (event-specific knowledge) than on the higher levels of organization (general event knowledge and lifetime periods). Berntsen and Rubin (2012), as well, propose a hierarchical structure along the different levels of abstractions on all dimensions. In order to further investigate the validity of those theories, future research could investigate the current networks for hierarchical clustering. Besides that, connecting to the three levels of abstraction along the temporal dimension as posited by Conway and Pleydell-Pearce (2000), future research could focus on examining a tripartite network that explicitly assesses a higher level of organization. Subsequently, similar procedures to what was done in this paper could follow.
In addition, future research could use the method that was developed in this paper to
investigate the posited flexible nature of autobiographical memory structure as proposed by Berntsen and Rubin (2012). For example, figure 4 presents the structure defined solely by people. We can observe that it is a sparser network with much more clustering going on. It illustrates how this
“flexible structure” approach would alter what memories are connected. In order to further
investigate the validity of the theory by Berntsen and Rubin (2012), networks where memories are connected based on a shared period of time, location and people could be constructed and analyzed.
Figure 4. Memory network of participant 1, based solely on people. The darker the node the higher its clustering coefficient. Layout by Fruchterman & Reingold (1991).
Despite its implications, this study knows conceptual and methodological limitations. During modification of the pulled features we did not manage stemming and pluralization, like pulling “help” when the participant wrote “helped” or “house” when the participant wrote “houses.” Therefore, we may have missed connections between memories. Besides that, participants recalled only 100 memories – a small percentage of the total number of memories we have. We may also have gathered a biased sample of memories that does not represent the full autobiographical memory structure, as the gathered memories would possibly prove to be quite central in the most complete
possible autobiographical memory network. Furthermore, our small-world index did not incorporate the weights of the edges, i.e. it did not distinguish between memories that share one feature and memories that share 10 features. By doing so, in future research, we could provide a more nuanced view which could especially be of interest for the memory network.
Those limitations let us place the study in a broader context: we are left with the question what we can actually regard as salient features of autobiographical memories. Although we can assume that well-known features like the temporal, spatial and social aspects do play a role in organization of autobiographical memories (Wagenaar, 1986; Berntsen & Rubin, 2012), there might be other important features or dimensions that have not been given attention in the current study, e.g. emotional intensity (Talarico, LaBar & Rubin, 2004). To understand what kind of features play a role in memory organization and retrieval, further research could focus on the nature of different features and how salient they are in autobiographical memory organization.
Conclusion
All in all, this study is an addition to the field of autobiographical memories, as this is the first time that autobiographical memories have been mapped out as a network. The limitations do not severely disqualify the found results. Instead, they can pave the way for future research. The found small-worldness in the feature network indicates that
autobiographical memory features obtain a certain structure that may allow for effective storage and retrieval. That we did not find small-worldness in the memory network could either mean that there is no small-worldness in the memory network or that there is another underlying structure. The current study helps to fill in the larger picture of autobiographical memory storage and retrieval – something that each of us engages in in everyday life.
References Auxiliary verbs. (2018). Retrieved from
http://www.myenglishpages.com/site_php_files/grammar-lesson-auxiliary-verbs.php Barabási, A. L. (2011). The network takeover. Nature Physics, 8(1), 14.
Barabási, A. L. (2012). Network Science Book Project. Network Science Book Project. Bassett, D. S., & Bullmore, E. D. (2006). Small-world brain networks. The
Neuroscientist, 12(6), 512-523.
Benoit, K., Muhr, D., & Watanabe, K. (2017). Stopwords: multilingual stopword lists. Retrieved from https://CRAN.R-project.org/package=stopwords
Berntsen, D., & Rubin, D. C. (2004). Cultural life scripts structure recall from autobiographical memory. Memory & Cognition, 32(3), 427-442.
Berntsen, D., & Rubin, D. C. (2012). Understanding autobiographical memory: an ecological theory. In D. Berntsen & D. C. Rubin (Eds.), Understanding
autobiographical memory: Theories and approaches (pp. 333-355). Cambridge:
Cambridge University Press
Brunec, I. K., Chadwick, M. J., Javadi, A. H., Guo, L., Malcolm, C. P., & Spiers, H. J. (2015). Chronologically organized structure in autobiographical memory search. Frontiers in Psychology, 6(2), 338.
Borsboom, D., Cramer, A. O., Schmittmann, V. D., Epskamp, S., & Waldorp, L. J. (2011). The small world of psychopathology. PLOS One, 6(11), e27407.
Conway, M. A. (2005). Memory and the self. Journal of Memory and Language, 53(4), 594-628.
Conway, M. A., & Pleydell-Pearce, C. W. (2000). The construction of autobiographical memories in the self-memory system. Psychological Review, 107(2), 261-288.
Eade, J., Healy, H., Williams, J. M. G., Chan, S., Crane, C., & Barnhofer, T. (2006). Retrieval of autobiographical memories: The mechanisms and consequences of truncated search. Cognition and Emotion, 20(3-4), 351-382.
Essberger, J. (2012). English prepositions list. EnglishClub.com. Retrieved from
https://s3.amazonaws.com/academia.edu.documents/36666068/EnglishClub-English-Prepositions-List.pdf?AWSAccessKeyId=AKIAIWOWYYGZ2Y53UL3A&Expires =1527858333&Signature=OafpLnrCJm38sb32WuOQmDw6MWU%3D&response content-disposition=inline%3B%20filename%3DEnglish_Club_English_Prepositions _List.pdf
Fruchterman, T. M., & Reingold, E. M. (1991). Graph drawing by force‐directed placement. Software: Practice and Experience, 21(11), 1129-1164.
Greenberg, D. L., & Verfaellie, M. (2010). Interdependence of episodic and semantic memory: evidence from neuropsychology. Journal of the International
Neuropsychological Society, 16(5), 748-753.
Haque, S., Juliana, E., Khan, R., & Hasking, P. (2014). Autobiographical memory and hierarchical search strategies in depressed and non-depressed participants. BMC
Psychiatry, 14(1), 310-319.
Hills, T. T., Jones, M. N., & Todd, P. M. (2012). Optimal foraging in semantic memory. Psychological Review, 119(2), 431-437.
Humphries, M. D., & Gurney, K. (2008). Network ‘small-world-ness’: a quantitative method for determining canonical network equivalence. PLOS One, 3(4), e0002051.
Ingo Feinerer and Kurt Hornik (2017). tm: Text mining package. R package version 0.7 3. Retrieved from https://CRAN.R-project.org/package=tm
Lists of Adjectives. (2018). Retrieved from
Lists of Pronouns in English. (2018). Retrieved from
http://www.really-learn-english.com/list-of-pronouns.html
Morais, A. S., Olsson, H., & Schooler, L. J. (2013). Mapping the structure of semantic memory. Cognitive Science, 37(1), 125-145.
Mullen, L., Keyes, O., Selivanov, D., Arnold, J., & Benoit, K. (2018). Tokenizers: Fast, consistent, tokenization of natural language text. Retrieved from https://cran.r project.org/web/packages/tokenizers/index.html
Neisser, U. (1986). Nested structure in autobiographical memory.
Qualtrics [Computer software]. (2018). Retrieved from https://www.qualtrics.com/ Rinker, T. (2018). qdapDictionaries: Dictionaries and word lists for the ‘qdap’ package.
Retrieved from https://CRAN.R-project.org/package=qdapDictionaries
RStudio (2018). RStudio: Integrated development environment for R (Version 1.1.442) [Computer software]. Retrieved from https://www.rstudio.com/
Singer, J. A., & Moffitt, K. H. (1992). An experimental investigation of specificity and generality in memory narratives. Imagination, Cognition and Personality, 11(3), 233-257.
Talarico, J. M., LaBar, K. S., & Rubin, D. C. (2004). Emotional intensity predicts autobiographical memory experience. Memory & Cognition, 32(7), 1118-1132. Telesford, Q. K., Joyce, K. E., Hayasaka, S., Burdette, J. H., & Laurienti, P. J. (2011). The
ubiquity of small-world networks. Brain Connectivity, 1(5), 367-375. Thorne, A., & McLean, K. C. (2001). Manual for coding events in self-defining
memories. Unpublished Manuscript, University of California, Santa Cruz.
Tulving, E. (1972). Episodic and semantic memory. Organization of Memory, 1(1), 381-403. Tulving, E., & Murray, D. (1985). Elements of episodic memory. Canadian
Wagenaar, W. A. (1986). My memory: A study of autobiographical memory over six years. Cognitive Psychology, 18(2), 225-252.
Watts, D. J., & Strogatz, S. H. (1998). Collective dynamics of ‘small-world’ networks. Nature, 393(6684), 440-442.
Williams, H. L., Conway, M. A., & Cohen, G. (2008). Autobiographical memory. Memory in
